Extracting Rules from Autonomous-Vehicle-Involved Crashes by Applying Decision Tree and Association Rule Methods

Abstract

Autonomous vehicles (AVs) can dramatically reduce the number of traffic crashes and associated fatalities by eliminating the avoidable human-error-related crash contributing factors. Many companies have been conducting pilot tests on public roads in several states in the U.S. and other countries to accelerate AV mass deployment. AV pilot operations on Californian public roads saw 251 AV-involved crashes (as of February 2020). These AV-involved crashes provide a unique opportunity to investigate AV crash risks in the mixed traffic environment. This study collected the AV crash reports from the California Department of Motor Vehicles and applied the decision tree and association rule methods to extract the pre-crash rules of AV-involved crashes. Extracted rules revealed that the most frequent types of AV crashes were rear-end crashes and predominantly occurred at intersections when AVs were stopped and engaged in the autonomous mode. AV and non-AV manufacturers and transportation agencies can use the findings of this study to minimize AV-related crashes. AV companies could install a distinct signal/display to inform the operational mode of the AVs (i.e., autonomous or non-autonomous) to human drivers around them. Moreover, the automatic emergency braking system in non-AVs could avoid a significant number of rear-end crashes as, often, rear-end crashes occurred as a result of the failure of following non-AVs to slow down in time behind AVs. Transportation agencies can consider separating AVs from non-AVs by assigning “AV Only” lanes to eliminate the excessive rear-end crashes resulting from the mistakes of human drivers in non-AVs at intersections.

Traffic crash-related fatalities have been one of the major causes of death in the U.S. and around the world. In 2018, 37,000 died on the U.S. roadways as a result of traffic crashes, and this caused an economic loss of $147 billion ( 1 ). Transportation agencies in the U.S. have developed “Vision Zero” initiatives to eliminate all traffic fatalities and severe injuries ( 2 ). Over 90% of all crashes happen because of driver errors, and over 40% of fatal crashes happen as a result of some combination of alcohol, distraction, drug involvement, and/or fatigue. As most crashes occur because of driver errors, emerging autonomous vehicles (AVs) have the potential to eliminate driver errors and associated traffic crashes and help agencies to achieve the Vision Zero goal. Besides, AVs can reduce congestion and fuel consumption and can improve mobility performance by reducing vehicle ownership and providing mobility services to elderly and disabled people ( 2 ). Since September 2014, the California Department of Motor Vehicles (CA DMV) allowed permit-holding companies to test AV technologies on the public roadways. Human test/safety drivers in these AVs are responsible for disengaging from autonomous mode to manual mode when the autonomous mode faces any technical issues or approaches an unfamiliar driving environment. While simple/typical driving actions (e.g., maintaining position in a lane) are relatively easy for an AV to operate, designing an autonomous system that can perform safely in nearly every situation is challenging ( 3 ). Navigating complex situations, such as intersections, where the number of potential conflict points is high, poses a great challenge for AVs. Testing companies operating in CA are required to submit a Traffic Collision Involving an Autonomous Vehicle Report (OL 316 Form) to CA DMV. From September 2014 to February 2020, 251 AV-related crashes occurred in California (CA).

Ensuring the safety of AV technology development is a complex challenge, and improvement of AV technology depends on addressing the wide array of risk factors, such as the behavior of non-AV drivers, driving environments, and AV technology limitations. To minimize the crash risk of AVs, AV technology has to detect and predict unsafe movements of other vehicles and roadway users in complex driving environments. CA AV crash reports provide a unique opportunity to identify the challenges faced by AVs during the pilot operation on public roads and develop potential solutions for AV technology improvement. This study compiled the CA AV crash reports to analyze the crash contributing factors that led to rear-end, sideswipe, broadside, or head-on types of AV-involved crashes. The results of this study developed new insights on the safety performance of AVs in real-world driving environments and potential measures that can be implemented by different AV stakeholders to reduce AV safety deficiencies.

The structure of the paper is as follows: the literature review section provides a comprehensive overview of the past studies focused on AV crashes, followed by the gap in the existing literature and how this study addresses the gap. The next four sections explain the AV-related crash data, adopted research method for the crash rules extraction, the results of the rules extraction, and potential implications of the findings. The last section presents concluding remarks.

Literature Review

Since the beginning of testing on public roadways, AV technology has made significant progress. An AV uses a combination of hardware and software technologies, such as cameras, global positioning systems (GPS), radio detection and ranging (RADAR) sensors, light detection and ranging (LiDAR) sensors, digital maps, video cameras, and computing platforms to sense the driving environment and determine AV movement actions. The number of companies conducting AV testing on public roads has been increasing, and many of the tests have been performed in a closed environment (i.e., test tracks/proving grounds) ( 4 ). Many companies and researchers have been using driving simulators as a tool for AV technology development. Several studies investigated the driver’s behavior and performance in a highly automated driving system ( 5 – 7 ). Complex driving environments could lead to a manual takeover of the AV by the safety driver, which is known as “disengagement” of AV technology. In addition to the disengagement reports, CA DMV maintains a separate database for all AV-involved crash reports since 2014. AV crash reports document some important crash information, such as manufacturers’ information, crash information, serious injuries to people, and other associated factors such as weather, lighting, and pavement condition ( 4 ). As of February 2020, the DMV has received 251 autonomous vehicle collision reports. Prediction and estimation of traffic crashes and crash contributing factors have been an important research topic for decades. The traditional approach for non-AV crash modeling is to develop statistical models to explore the underlying causes of the crashes ( 8 – 10 ). Besides these statistical modeling approaches, researchers also used different machine learning (ML) and neural network methods to analyze traffic crashes ( 11 – 14 ). Many studies applied non-parametric methods, such as decision trees (DT) or association rules to identify the crash contributing factors. For example, Montella et al. used the classification and regression tree (CART) model and association rules to generate rules for powered two-wheeler crashes in Italy ( 15 ). De Oña et al. and Abellán et al. used tree-based methods to analyze crash severity and crash types ( 16 , 17 ).

The availability of AV crash databases is very limited, and only CA DMV has released AV crash data to the public. Few researchers have explored the mechanism of AV crashes using the CA AV crash data. Previous research on AV disengagements and crashes focused on the exploratory analysis of the crash data to find the relation between different contributing factors. Favaro et al. examined the factors of AV crashes by exploring the drivers’ engagement at pre- and post-crash scenarios, the relative speed of the involved vehicles, and the location of the crashes ( 18 ). Boggs et al. conducted a statistical analysis of the CA AV rear-end type of crashes and injury crashes using text mining techniques and a hierarchical Bayesian-heterogeneity-based approach ( 19 ). The chance of a rear-end crash was higher in situations when the AV is driven in an autonomous mode compared with the situations when the AV was driven in non-autonomous mode. Wang and Li explored the AV crash severity and crash types by applying the ordinal logistic and CART classification tree models ( 4 ). Several studies also explored the AV disengagement data. Favarò et al. analyzed the disengagement triggers, trends, and contributory factors, and found that 1 in every 178 disengagements resulted in a crash ( 20 ).

A key gap in the limited AV-related crash studies is the comprehensive analysis of the AV and non-AV movements in the mixed traffic conditions that resulted in different types of AV-involved crashes. This study investigated the AV and non-AV movements, as well as other pre-crash factors (e.g., driving mode, location of the crash) in AV-involved rear-end, sideswipe, broadside, and head-on crashes. Previous literature on non-AV crashes suggests that analyzing crash types has three primary benefits: (1) identification of locations deficiency to specific crash types which are not likely to be revealed by total crash model; (2) implementation of the improved and efficient type of countermeasures, as there are target crashes revealed through crash types analysis; and (3) better development of crash predictive models, as the association of individual crash types with various covariates such as roadway geometry, traffic volume, and environmental factors ( 21 – 25 ). Thus, in this study, the DT and association rule methods were applied to generate and identify major combinations of crash contributing factors in AV-involved crashes based on model performance criteria.

CA DMV AV Crash Data

Most of the AV technology companies in CA operated their test AVs in the San Francisco and Santa Clara areas. A total of 198 AV-related crashes that occurred between January 2016 and February 2020 were used in this study. A crash database was developed for this study using the crash reports. A summary statistic of different characteristics of the AV crashes is provided in Table 1.

Table 1.

Summary Statistics of the Autonomous Vehicle (AV) Crashes and Explanatory Variables

Variable	Category	No.	Percentage (%)	Variable	Category	No.	Percentage (%)
County				Lighting condition
	Alameda	2	1.0		Dark/street light	30	15.2
	Contra Costa	1	0.5		Daylight	121	61.1
	San Francisco	124	62.6		Dusk/dawn	5	2.5
	Santa Clara	71	35.9		NA	42	21.2
Crash type				Crash severity
	Broad Side	15	7.6		None	19	9.6
	Head-on	17	8.6		Moderate	21	10.6
	Sideswipe	40	20.20		Minor	135	68.2
	Rear-end	122	61.6		Major	1	0.5
	Other	4	2	AV movement
					Stopped	78	39.4
Operational mode					Slowing/stopping	22	11.1
	Non-autonomous	55	28		Proceeding straight	75	37.9
	Autonomous	114	57		Turning movement	17	8.6
	Disengaged	29	15	Non-AV movement
Location of the crash					Stopped	6	3.0
	Ramp	5	2.5		Slowing/stopping	14	7.1
	Intersection	126	63.6		Proceeding straight	128	64.6
	In roadway	61	30.8		Turning movement	31	15.7
	Parking lot	6	3.0

Note: NA = not available.

Crash Types and AV Movement

Four main types of AV-related crashes (rear-end, sideswipe, broadside, and head-on) were considered in this study. Non-AVs hitting the rear-side of the AVs were defined as rear-end crashes. The summary statistics in Table 1 shows that about 62% of the AV crashes were rear-end crashes. Most of these rear-end crashes happened when the AVs were stopped or slowing down to a full stop at intersections. The second most common AV crash type was the sideswipe crash. Sideswipe crashes were defined as the crashes where the non-AV struck the side of the AV, and vice versa. A total of 40% of the sideswipe crashes happened when the AV and non-AV were moving in the same direction (Table 1). Broadside crashes were defined as crashes when the side of the AV was struck by the non-AV at an angle. A total of 15 crashes in the crash database were broadside crashes, which was only 7.6% of the total number of AV-involved crashes (Table 1). Broadside crashes occurred when an AV was traveling straight at an intersection and hit by a non-AV making a left-turn or right-turn. Lastly, “head-on” was the type of crash when the front of an AV was struck by a non-AV, or the AV struck another vehicle or object. In the AV crash database, only 8.6% of the total crashes were head-on crashes (Table 1). The AV movement data in Table 1 shows that the most frequent AV movements during the crashes were “AV proceeding straight” and “AV stopped.” In comparison, the predominant non-AV movements were “proceeding straight” (64.6%) and “turning movements” (15.7%), as, in most of the cases, AVs were hit by moving non-AVs (Table 1).

Crash Severity

CA AV crash reports classify crash severity (vehicle damage) into four types—Major, Minor, Moderate, and None. In most frequent rear-end crashes, the relative speed of the AV and non-AV were very low (i.e., less than 10 mph). As a result, most of the rear-end crashes caused minor damage to the AVs, except for one rear-end crash with major damage to the AV. A total of 21 out of 198 AV crashes experienced moderate damage to the AVs, which was only 10.6% of total crashes (Table 1).

AV Driving Mode

AV manufacturers are required to report the driving mode (i.e., autonomous or non-autonomous mode) for each AV crash event. Crash reports showed that 57% of the crashes occurred when the AV was operating in autonomous mode. In some instances, AVs were disengaged to non-autonomous mode right before the crashes. These disengagements generally happened when the autonomous technology failed or when the safety driver identified an issue and took over the AV operation. A total of 15% of the AV crashes occurred right after the autonomous mode was disengaged, and 28% of the crashes occurred when the AVs were driving in non-autonomous/conventional mode (Table 1).

Methods: Rules Extraction

Each AV-involved crash was the result of a combination of crash contributing factors, and each crash can be expressed as a logical combination of crash contributing factors. These conditional statements are called crash rules. In this study, rules were extracted by applying two different techniques—DT and association rule methods. Both methods are non-parametric and provide complementary insights on AV-involved crashes. The non-parametric method such as DT and association rules are better suited for this study compared with other parametric statistics, as the AV crash database has a relatively small sample size and the CA DMV AV crash database provided only a few factors associated with these crashes. The results from the DT and association rules provide intuitive and direct correlation factors that could lead to the appropriate mitigation measures to improve the safety performance of AVs during the critical ongoing technology development phase. Thus, the rules generation by DT and association rules identified a set of important rules in the AV crash database, which can be used by AV safety analysts to improve the safety of the AV technology.

In the DT method, each terminal node of the tree represents a rule, with all the preceding nodes that are connected to the terminal node acting as the crash contributing factors. Each terminal node of the tree gives the probability of each crash type. On the other hand, association rules are based on the relative frequency of the combination of factors that occurred in the crash database. The rules can have crash types, such as rear-end, sideswipe, broadside, or head-on, as consequent. The explanatory variables used for the rule extraction were the driving mode of the AVs, disengagement of the autonomous mode, location of the crash, and movement of the AVs and non-AVs at the moment of the crash.

Decision Tree (DT) Model

DT is a non-parametric predictive model that does not require prior probabilistic knowledge on the study phenomena ( 16 ). DT can be used for classification and regression modeling purposes. If the dependent variable in the DT is a class variable (e.g., crash type) and has a limited number of possible options, then it is called classification. DT represents a simplistic and graphically hierarchical structure and is used to extract rules and to identify the effect of different explanatory variables (i.e., crash contributing factors). In the hierarchical structure of a DT, each node represents a feature, and each branch represents one of the conditions of that feature variable. The terminal node of the tree, known as a tree leaf, represents the expected value of the class variable. Each node is associated with the most informative variable that has not already been selected in the path from the root to that node.

Several algorithms can be used to build a DT. The splitting criteria of the tree formation are the main difference between these methods ( 16 ). Among all algorithms, the CART model developed by Breiman et al. is the most widely used method in traffic crash data analysis ( 26 ). In this study, the CART algorithm was used to build the DT in the JMP Pro program. The CART model uses the Gini index as splitting criteria, which measures the degree of impurity at each node, until their purity cannot be increased by splitting further. The Gini index for a variable C (for example operational mode of the AVs: autonomous or non-autonomous) can be defined as the following Equation 1:

gini (C) = 1 - \sum_{j} p^{2} (C); C = c_{j}

(1)

where

p is the probability of a variable C being classified to a particular class c_j at a node of the DT (e.g., for the variable “driving mode,”p is the probability of being the driving mode autonomous or non-autonomous at that node).

The split criteria based on the Gini index can be defined as Equation 2:

GIx (C, X) = gini (C | X) - gini (X)

(2)

where

$gini (C | X) = \sum tP (X_{t}) gini (C | X = x_{t})$ ; and

X is another known independent variable (for example, the location of the AV crashes).

The best split minimizes GIx (C, X) ( 16 ). Similar to other classification methods, the performance of the CART model is measured by the accuracy, which is defined as the percentage of cases correctly classified by the classifier.

A decision rule is a logical conditional expression that states the sequence of the events that lead to the occurrence of a particular class. It takes the form “IF A THEN B” where A is the antecedent, and B is the consequent. The rules generated from a DT depend on the direction of the variable at the root node, where the conditional “IF” structure begins. To separate the rules that are significant, three parameters (i.e., support [S], population [Po], and Probability [P]) were used in this study. The support is the percentage of the dataset where A and B occur, the population is the percentage of the dataset where A occurs, and Probability P is the percentage where the rule is accurate. The minimum value of these parameters depends on the nature of the dataset (balanced or unbalanced), rare events, and the size of the sample. For example, Montella et al. used a low value of probability (P) because their data was unbalanced, whereas De Oña et al. used a high value of P as their data was balanced ( 15 , 16 ). As the data used in this study contains a limited number of crash events, and the data is unbalanced, a probability value of 40% as a minimum value was used to separate the significant rules.

Association Rule Data Mining

Association rule is one of the most popular non-parametric data mining techniques that can be used in traffic safety causality analysis. The main advantage of the association rule data mining method is that the dependent variable does not need to follow a specific functional distribution to generate the rules. Generated association rules can be used to develop countermeasures to break the association between the contributing factors to reduce the crash risk ( 27 ). The association rule discovery is usually done in two steps. In the first step, factors that are frequent (e.g., AV movements, locations) and have a support value greater than the threshold minimum value are selected. Factors are divided into two types—antecedent and consequent. From the list of contributing factors in a crash, one factor is set as consequent, and others are set as antecedents and are expressed as conditional forms called rules. Then, the confidence value of the rules (generated in the first step) is used to determine the significance of the rules. Next, one item is subtracted from the antecedent and added to the consequent to check the confidence value of the new rules. This process is iterated until the antecedent becomes empty. The second step of the association rule discovery method is more straightforward. In this step, rules generated in the first step are organized based on their lift values ( 28 ).

Several algorithms can be used for developing association rules. Apriori and Eclat are the two most commonly used algorithms by researchers over the years. The Apriori algorithm was first introduced by Agarwal et al. ( 29 ). In this study, the Apriori algorithm is used for mining the association rules from the CA DMV AV crash database using the program R. The working principle of this algorithm is discussed below:

Let, $I = {i_{1}, i_{2}, i_{2} \dots \dots ., i_{n}}$ be a set of n attributes called items such as driving mode, lighting condition, or vehicle movement type, and $D = {c_{1}, c_{2}, c_{2} \dots \dots ., c_{m}}$ contains m numbers of AV crash causality data. Each crash event in D occurred as a result of some unique combination of attributes from I and has a unique id (i.e., crash ID). So, each causality in D contains a subset of the items in I. A rule is defined as an implication of the form X⇒Y where X, Y⊆I, and X∩Y = ∅. The part of an association rule (X) is antecedent(s), while the second part (Y) is consequent. In this study, crash types (i.e., rear-end crash, sideswipe crash, broadside, and head-on) were considered as consequent, and different crash contributing factors, such as the location of crash and driving mode, were considered as antecedents. To separate significant rules from the whole set of rules generated by the Apriori algorithm, various measures of interest and significance can be used. The best-known measures are minimum thresholds on support and confidence ( 30 ). The support of a rule is the percentage of the entire dataset covered by the rule and expressed as Equation 3:

Support (X \to Y) = \frac{# (X \cap Y)}{N}

(3)

where

N is the total number of AV crash events.

The confidence of a rule is the conditional probability of occurrence of the consequent, given that the antecedent(s) have occurred, and is expressed as Equation 4:

Confidence = \frac{Support (X \to Y)}{Support (X)}

(4)

In addition, the lift value was used to measure the degree of interdependence among antecedent and consequent of the rule. The lift measures the ratio of confidence of a rule and the expected confidence that the consequent occurs depending on the occurrence of the antecedents, expressed as Equation 5:

Lift = \frac{Support (X \to Y)}{Support (X) \times Support (Y)}

(5)

The lift value shows the frequency of the occurrence of antecedent and consequent ( 27 ). Lift value of a rule less than one indicates negative interdependence between the antecedent and the consequent, and greater than one indicates that the antecedent and the consequents are positively interrelated. If the lift value is equal to one, then the antecedent and consequents are independent ( 31 ). A large confidence, and a lift value greater than one are desirable for the rules to have a high level of support. Moreover, lift value is more important for determining the strength of an association rule than support and confidence.

Results and Discussion

Results of the DT Model

The DT was developed using the CART model, where 70% of the crash events were used for training the model, and 30% of the data were used to validate the model. The accuracy of the CART model using the training dataset and validation dataset was about 69% and 65%, respectively. The observed model accuracy was similar to the previous studies that focused on crash severity modeling and identified the associated crash contributing factors, where the focus of this study was on crash type modeling. De Oña et al. used the DT method to extract rules from the police accident reports, where the accuracy of the CART model was 55.87% ( 16 ). Abdelwahab and Abdel-Aty, and De Oña et al. also reported similar accuracy on their crash severity modeling studies using data mining and Bayesian networks, respectively ( 32 , 33 ). Figure 1 shows the developed DT for AV-involved crashes compiled for this study.

Figure 1.

Decision tree (DT) of autonomous vehicle (AV)-involved crashes, crash types as response variable.

“Non-AV movement” was the root node that generated the DT, which then divided into two branch nodes (node 1 and node 2), as shown in Figure 1. Each node was then further divided into branch nodes based on the AV movement, driving modes, and location of the crashes. There was a total of 26 nodes in the DT developed using the CART model. The terminal node represents the probability of each crash type as a combination of factors that appear on the previous branch nodes leading to the terminal node. The decision rules can be expressed as a conditional statement of the nodes in different levels in the DT. A total of 14 rules (in relation to crash probability) were generated from the DT, and the top 10 significant rules are listed in Table 2. Each rule estimated the crash risk of different crash types for a combination of crash contributing factors.

Table 2.

Extracted Rules from the Decision Tree (DT) Model

Rule no.	Leaf label	Crash type probability
Rule no.	Leaf label	Broadside	Head-on	Rear-end	Sideswipe
1	Non-AV movement (non-AV PS, non-AV slowing) and location (in roadway) and driving mode (non-autonomous) and AV movement (AV stopped)	0.009	0.156	0.802	0.033
2	Non-AV movement (non-AV PS, non-AV slowing) and location (in roadway) and driving mode (non-autonomous) and AV movement (AV PS)	0.006	0.099	0.510	0.385
3	Non-AV movement (non-AV PS, non-AV slowing) and location (ramp, intersection) and driving mode (autonomous) and AV movement (AV PS)	0.004	0.004	0.626	0.366
4	Non-AV movement (non-AV PS, non-AV slowing) and location (ramp, intersection) and driving mode (non-autonomous) and disengagement (no) and AV movement (AV slowing, AV stopped)	0.013	0.008	0.828	0.151
5	Non-AV movement (non-AV TM, non-AV stopped) and disengagement (no) and location (in roadway) and AV movement (AV PS)	0.013	0.864	0.090	0.033
6	Non-AV movement (non-AV PS, non-AV slowing) and location (ramp, intersection) and driving mode (non-autonomous) and disengagement (no) and AV movement (AV TM, PS)	0.518	0.011	0.437	0.035
7	Non-AV movement (non-AV TM, non-AV stopped) and disengagement (no) and location (intersection) and driving mode (autonomous) and AV movement (AV PS, AV stopped)	0.007	0.103	0.870	0.020
8	Non-AV movement (non-AV TM, non-AV stopped) and disengagement (no) and location (in roadway) and AV movement (AV TM, AV stopped, AV slowing)	0.154	0.312	0.506	0.029
9	Non-AV movement (non-AV TM, non-AV stopped) and disengagement (no) and location (intersection) and driving mode (autonomous) and AV movement (AV TM)	0.153	0.019	0.367	0.460
10	Non-AV movement (non-AV TM, non-AV stopped) and disengagement (no) and location (intersection) and driving mode (non-autonomous)	0.007	0.414	0.156	0.423

Note : AV = autonomous vehicle; PS = proceeding straight; TM = turning movements.

Rules 1, 2, 3, 4, 7, and 8 (six rules out of the top 10 rules) in Table 2 identified the higher probability of rear-end crashes. Three out of these six rules (rules 3, 4, and 7) had intersections as the location of crashes which means that rear-end AV crashes were more likely to happen at intersections. According to the pre-crash factors of DT rule 1, when an AV stopped at a roadway and a non-AV was slowing or proceeding straight behind the AV, there was 80% chance of rear-end crash for non-autonomous mode of driving. DT rule 2 indicates that the probability of rear-end crash was also high (51%) when the AV was proceeding straight on a roadway in a non-autonomous driving mode and the non-AV was slowing down or proceeding straight behind the AV. Pre-crash factors of DT rules 3 and 4 show that the rear-end AV crashes also occurred when the AVs were traveling at the ramps or intersections. DT rule 7 had the highest probability (87%) of rear-end crash when the AV was stopped in autonomous mode and the non-AV was making a left- or right-turn movement at an intersection. Most of the rear-end crashes occurred because of unsafe movement of human-driven non-AVs as they failed to stop before hitting the back of the AVs.

The second most common type of AV-related crash was sideswipe crashes. Sideswipe crashes were more likely to happen when both the AV and non-AV were proceeding forward or making turning movements in roadways or at intersections. DT rules 1 and 2 show that, when the location of the crash was on a roadway (i.e., not at an intersection) and the AV was proceeding forward (rule 2) instead of stopped at an intersection (DT rule 1) the chance of sideswipe crashes increased from 9% (DT rule 1) to 38% (DT rule 2). Chances of sideswipe crash were 46% and 42% according to DT rules 9 and 10, respectively, where crashes happened when the AV and non-AV were making turning movements at the intersection. Sideswipe crashes occurred during the AVs’ non-autonomous and autonomous driving mode. Rule 2 showed that sideswipe crashes occurred when both the AV and non-AV were proceeding straight in a roadway. Most of the sideswipe crashes happened because of the fault of non-AVs. At an intersection, sideswipe crashes occurred when both the AV and non-AV were making turning movements, while the non-AV crossed its lane and contacted the AV. In other cases where sideswipe crashes occurred in a roadway, the non-AV sideswiped the AV while passing and merged back into the AV’s lane too closely.

On the other hand, DT rule 5 indicates 86% chance of a head-on crash when the AV was proceeding straight, and the non-AV was stopped or making a left- or right-turning movement. According to the rule, head-on crashes were more likely to happen at locations such as the middle or rightmost lanes and right-turning lanes, rather than intersections. Rule 10 also showed that the chance of head-on crash was 42% when the location of the crash was an intersection, the AV was in non-autonomous mode, and the non-AV was performing a turning movement or stopped. Analysis of the movements of AVs and non-AVs from these two rules (rules 5 and 10) showed that head-on crashes occurred mainly because of the non-AV. According to these rules, the non-AV traveled in the wrong direction in a one-way road after making a left-/right-turn or aggressively swerved lanes and stopped in front of an AV. A few of the head-on crashes also occurred because of the AV, while it was running in the non-autonomous mode (rule 10). Finally, rule 6 (in Table 2) describes the combination of factors that created a scenario where the chance of broadside crash was 51.8%, and the chance of rear-end crash was 43.7%. This rule showed that the predominant movement of the AV and non-AV at the moment of crashes was “proceeding straight,” and the location of the crash was “intersection.” Further investigating the crash reports, it was found that almost all broadside crashes occurred because of the non-AVs, where the non-AV either disobeyed the stop sign at the intersection or violated the red light and proceeded through the intersection.

Overall, the rules indicated that AVs performed well in preventing sideswipe, broadside, and head-on crashes, meaning that the rear-end crashes where the AV was in front of the non-AV was the most critical crash scenario to be addressed. AVs are designed to follow the traffic rules and maintain posted speed limits, where human drivers in non-AVs choose a safe and comfortable speed, often higher than the posted speed limit. Thus, the speed differential between AVs and non-AVs could lead to rear-end crashes. Intersections in mixed traffic conditions (i.e., a mix of AVs and non-AVs) presented a challenging environment for the AVs. Overcautious (e.g., longer startup delay by AVs at intersection stop bar) during the turning or forward movements at the intersection led the non-AV drivers (with shorter startup delay) to hit the AV and cause a rear-end crash, or to pass the AV and cause a sideswipe crash.

Rules Generated from Association Rules

For the association rules generation, six crash attributes (location of the AV crash, mode of driving, disengagement, AV movement during the crash, non-AV movement during the crash, and crash type) were used. As rear-end and sideswipe were the most frequent types of AV crash, and the other two types of crashes (i.e., broadside, head-on) were rare in the database, a low value of support and confidence was used for the generation of the rules ( 34 ). In this study, by trial and error, the minimum support value of 0.03 and confidence value of 0.4 were selected, and the total number of rules generated after the initial training of the algorithm was 6,486. The lift value of all generated rules was greater than one, which means that the antecedent and consequents were positively interdependent in all rules. Next, the rules with crash types as the antecedent were separated to find the contributing factors behind the different types of crashes. The extracted rules from the association rule method using crash types as an antecedent is presented in Table 3.

Table 3.

Rules Generated from Association Rule Method Based on Crash Type

Rule no.	Antecedent	Consequent	Support	Confidence	Lift
1	Autonomous mode, intersection, AV stopped, non-AV slowing	Rear-end	0.050	1.000	1.631
2	Autonomous mode, intersection, AV slowing, non-AV proceeding straight	Rear-end	0.030	1.000	1.630
3	Autonomous mode, intersection, AV stopped, non-AV turning movement	Rear-end	0.035	1.000	1.630
4	Autonomous mode, intersection, AV stopped, non-AV turning movement, not disengaged	Rear-end	0.050	1.000	1.631
5	Autonomous mode, in roadway, AV proceeding straight	Sideswipe	0.030	0.462	2.296
6	Non-autonomous mode, intersection, AV turning movement	Sideswipe	0.035	0.438	2.177
7	Autonomous mode, AV proceeding straight, non-AV proceeding straight, not disengaged	Sideswipe	0.050	0.435	2.163
8	Non-autonomous mode, intersection, AV proceeding straight, non-AV turning movement	Broadside	0.030	1.000	13.266
9	AV proceeding straight, non-autonomous mode, in roadway, disengaged, non-AV slowing	Head-on	0.040	0.571	5.415
10	AV proceeding straight, non-autonomous mode, in roadway, disengaged, non-AV stopped	Head-on	0.035	0.538	5.100

Note : AV = autonomous vehicle.

Most of the rules generated by the Apriori algorithm supported the rules from the DT model, and also provided additional insights from the AV crashes. The top four rules shown in Table 3, representing a combination of crash contributing factors related to rear-end crashes, were similar to the findings of the DT model. These four rules related to rear-end crash identified autonomous driving mode as one of the crashes’ contributing factors, which means that AVs in autonomous mode had a rear-end crash risk higher than the non-autonomous driving mode. This finding indicates that fully engaged AVs were at high risk of rear-end crash in mixed traffic conditions, where AV-related rear-end crashes happened because of the following non-AV driver’s inability to understand the movement of the leader AV (e.g., whether AV was stopped or slowing down).

Association rules 5, 6, and 7 (Table 3) were related to AV sideswipe crashes. These three rules show similar findings revealed by the DT rules 2, 3, 9, and 10 (Table 2). However, the DT rules did not capture the influence of the AV’s operating mode on head-on crashes. Association rules 9 and 10 showed that the AV’s non-autonomous driving mode was associated with head-on crashes, and the AV was disengaged to non-autonomous mode right before the crash occurrence. This phenomenon could be attributed to potential issues related to AV technologies’ failure to take appropriate driving actions, and that the AV safety driver recognized the risk and attempted to take over vehicle control. The DT model explained that the broadside crashes happened when the AVs were making turning movements, and the non-AVs were clearing the intersection. According to association rule 8 (Table 3), broadside crashes also occurred when the non-AVs were making turning movements, and the AVs were clearing the intersections in non-autonomous driving mode. High lift values in a rule indicate strong positive dependence between the antecedents ( 22 ). As shown in Table 3, association rule 8 has a high lift value of 13.27, indicating strong positive association between the antecedent events. This high lift value means that the probability of a broadside crash was higher than for other types of crash when these antecedent events (i.e., AV proceeding straight, AV driving in non-autonomous mode, non-AV turning movement) were present at an intersection-related crash. Again, this scenario captured by rule 8 was not explained in the rules generated from the DT method. Overall, results from the DT method and association rules were complementary to each other, as both the methods revealed new information that could be used to develop countermeasures to reduce AV-involved crashes in the future.

Implications of the Research Findings

The findings of this study have several practical implications and can be used to develop collaborative AV technology development efforts among multiple stakeholders such as AV developers, traffic safety officials, and law enforcement agencies. In future, the roads will be shared by different levels of AV—from level 0 with no automation to level 5 with full automation. This mixed traffic poses challenges for both non-AVs and AVs. AV-related crashes mainly occurred when drivers of non-AVs failed to execute an appropriate driving task. The AV crash database showed that about 122 rear-end crashes occurred in CA between January 2016 and February 2020, which is about 62% of the total AV-involved crashes. About 63% of the 122 AV-involved rear-end crashes happened when the AVs were stopped or slowing down to a full stop at intersections. This statistic indicates that the AV-involved rear-end crashes occurred mostly because of the following non-AV’s failure to react appropriately to avoid a rear-end crash with the leading AV. As the AV technology is designed to obey traffic laws, the possible reason for the high number of rear-end crashes is that the non-AVs were unable to recognize the driving pattern of the AVs. Non-AV manufacturers could develop countermeasures to prevent these rear-end crashes. To avoid this type of crash, AV technology developers can modify AVs’ behavior to operate more like human drivers to minimize the consequence of non-AVs’ mistakes. In addition to obeying the traffic rules, AVs have to execute driving strategies that account for uncertainty about the movement of non-AVs. AVs can also make non-AVs aware of them by using distinct active signals about the status of autonomous driving mode (i.e., active or inactive). Also, AVs can provide warnings (via signals or signs) to non-AV drivers when they drive too close to the AVs. On the other hand, non-AVs can install the automatic emergency braking (AEB) system to avoid hitting the rear-ends of the AVs. Recently, the majority of automakers planned to equip all new passenger cars and all new trucks with a low-speed AEB system by September 1, 2022, and September 1, 2025, respectively ( 19 ). As most AV-related crashes mainly occurred at or near intersections, transportation agencies can modify the intersections to accommodate the safe movement of AVs (e.g., separate lanes for AVs depending on availability of right-of-way).

Conclusions

The future of significant improvement in traffic safety largely depends on the successful deployment of AV technologies. Although AVs have recorded a significant amount of vehicle miles in testing, AV technology is still maturing, and various companies have been fine-tuning their AV technology through extensive research programs, including testing on public roads. The AV crash reports provided insightful information on the unique nature of AV-related crashes. In this research, association rule and DT methods were used to identify major patterns in AV crash contributing factors.

The most significant rules identified the challenges of AVs sharing the roads with non-AVs. Almost 62% of the total crashes were rear-end crashes, and predominantly occurred at intersections. Results of the generated rules suggested that the rear-end crash was more prominent when the AV was engaged in autonomous mode. Most of the rear-end crashes occurred when the AV was stopped at the intersection, and the non-AV was proceeding straight or slowing down behind the AV at intersections. For other crash types, both autonomous mode and non-autonomous modes of AVs were frequent. Sideswipe crashes occurred both at intersections and on a roadway when the non-AV attempted to pass the AV or made turning movements. Further analysis of the AVs’ and non-AVs’ movement at the moment of crash showed that all types of AV-involved crashes occurred mainly because of non-AVs. Rear-end crashes occurred because of the non-AVs’ failure to react appropriately to avoid rear-end crashes with leading stopped AVs at intersections. On the other hand, broadside, sideswipe, and head-on crashes occurred because of the non-AVs’ unsafe movements, red-light running, or disobeying the stop sign at the intersection.

The findings indicate that rear-end crashes can be avoided by taking few simple precautionary measures such as separating AVs from non-AVs by using AV-only lanes, installing low-speed AEBs in non-AVs, or using signals/signs in the AVs letting non-AVs know about the AV driving modes. Currently, CA DMV collects limited information about AV-involved crashes. Although CA DMV modified the crash report (OL 316 form) over the years, crash severity levels of AV-involved crashes were not collected using the KABCO (K: fatal, A: serious injury, B: moderate injury, C: minor injury, and O: no injury) scale. A more comprehensive data collection of AV-involved crashes, including crash injury severity, driver’s cognitive state, roadway traffic and environmental conditions, degree of engagement of the autonomous system, and vehicle acceleration and deceleration rates, will enable a robust understanding of the factors influencing AV crashes, which can accelerate the development of effective safety remedies. Availability of a larger AV crash database with crash severity data scale in the future can be used to develop crash-frequency-based and crash severity models as an extension to this research. Results from the crash severity and crash frequency models can be used to improve roadway design and as well as vehicle design to increase AV-related traffic safety. Furthermore, AV technology is evolving rapidly with more testing on public roads, and the engagement of human drivers in non-AVs with AVs is expected to change over time. Because of the continuous improvement of AV technology and more widespread testing in public roads in the near future, a more comprehensive AV crash database could provide new perspectives about the AV crash mechanism that is partially or completely different from what was discovered using the AV crash database in this study.

Footnotes

Author Contributions

The authors confirm contribution to the paper as follows: study conception and design: M. T. Ashraf, K. Dey, data collection: M. T. Ashraf, K. Dey; analysis and interpretation of results: M. T. Ashraf, K. Dey; draft manuscript preparation: M. T. Ashraf, K. Dey, S. Mishra, M.T Rahman. All authors reviewed the results and approved the final version of the manuscript.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Kakan Dey

Sabyasachee Mishra

Md Tawhidur Rahman

References

Webb

C. N.

Motor Vehicle Traffic Crashes as a Leading Cause of Death in the United States, 2015. National Center for Statistics and Analysis, Washington, D.C., 2018.

Fagnant

D. J.

Kockelman

Preparing a Nation for Autonomous Vehicles: Opportunities, Barriers and Policy Recommendations. Transportation Research Part A: Policy and Practice, Vol. 77, 2015, pp. 167–181.

Campbell

Egerstedt

How

J. P.

Murray

R. M.

Autonomous Driving in Urban Environments: Approaches, Lessons and Challenges. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, Vol. 368, No. 1928, 2010, pp. 4649–4672.

Wang

Exploring the Mechanism of Crashes with Automated Vehicles Using Statistical Modeling Approaches. PLoS One, Vol. 14, No. 3, 2019, p. e0214550.

De Winter

J. C.

Happee

Martens

M. H.

Stanton

N. A.

Effects of Adaptive Cruise Control and Highly Automated Driving on Workload and Situation Awareness: A Review of the Empirical Evidence. Transportation Research Part F: Traffic Psychology and Behaviour, Vol. 27, 2014, pp. 196–217.

Körber

Gold

Lechner

Bengler

The Influence of Age on the Take-Over of Vehicle Control in Highly Automated Driving. Transportation Research Part F: Traffic Psychology and Behaviour, Vol. 39, 2016, pp. 19–32.

Merat

Jamson

A. H.

Lai

F. C.

Daly

Carsten

O. M.

Transition to Manual: Driver Behaviour When Resuming Control from a Highly Automated Vehicle. Transportation Research Part F: Traffic Psychology and Behaviour, Vol. 27, 2014, pp. 274–282.

Fountas

Anastasopoulos

P. C.

Abdel-Aty

Analysis of Accident Injury-Severities Using a Correlated Random Parameters Ordered Probit Approach with Time Variant Covariates. Analytic Methods in Accident Research, Vol. 18, 2018, pp. 57–68.

Iranitalab

Khattak

Comparison of Four Statistical and Machine Learning Methods for Crash Severity Prediction. Accident Analysis & Prevention, Vol. 108, 2017, pp. 27–36.

10.

Xie

Lord

Zhang

Predicting Motor Vehicle Collisions Using Bayesian Neural Network Models: An Empirical Analysis. Accident Analysis & Prevention, Vol. 39, No. 5, 2007, pp. 922–933.

11.

Mussone

Ferrari

Oneta

An Analysis of Urban Collisions Using an Artificial Intelligence Model. Accident Analysis & Prevention, Vol. 31, No. 6, 1999, pp. 705–718.

12.

Abdelwahab

H. T.

Abdel-Aty

M. A.

Artificial Neural Networks and Logit Models for Traffic Safety Analysis of Toll Plazas. Transportation Research Record: Journal of the Transportation Research Board, 2002. 1784: 115–125.

13.

Riviere

Lauret

Ramsamy

J. M.

Page

A Bayesian Neural Network Approach to Estimating the Energy Equivalent Speed. Accident Analysis & Prevention, Vol. 38, No. 2, 2006, pp. 248–259.

14.

Huang

Peng

Wang

Luo

Interactive Risk Analysis on Crash Injury Severity at a Mountainous Freeway with Tunnel Groups in China. Accident Analysis & Prevention, Vol. 111, 2018, pp. 56–62.

15.

Montella

Aria

D’Ambrosio

Mauriello

Analysis of Powered Two-Wheeler Crashes in Italy by Classification Trees and Rules Discovery. Accident Analysis & Prevention, Vol. 49, 2012, pp. 58–72.

16.

de Oña

López

Abellán

Extracting Decision Rules from Police Accident Reports through Decision Trees. Accident Analysis & Prevention, Vol. 50, 2013, pp. 1151–1160.

17.

Abellán

López

De OñA

Analysis of Traffic Accident Severity Using Decision Rules via Decision Trees. Expert Systems with Applications, Vol. 40, No. 15, 2013, pp. 6047–6054.

18.

Favarò

F. M.

Nader

Eurich

S. O.

Tripp

Varadaraju

Examining Accident Reports Involving Autonomous Vehicles in California. PLoS One, Vol. 12, No. 9, 2017, p. e0184952.

19.

Boggs

A. M.

Wali

Khattak

A. J.

Exploratory Analysis of Automated Vehicle Crashes in California: A Text Analytics & Hierarchical Bayesian Heterogeneity-Based Approach. Accident Analysis & Prevention, Vol. 135, 2020, p. 105354.

20.

Favarò

Eurich

Nader

Autonomous Vehicles’ Disengagements: Trends, Triggers, and Regulatory Limitations. Accident Analysis & Prevention, Vol. 110, 2018, pp. 136–148.

21.

Sarker

A. A.

Paleti

Mishra

Golias

M. M.

Freeze

P. B.

Prediction of Secondary Crash Frequency on Highway Networks. Accident Analysis & Prevention, Vol. 98, 2017, pp. 108–117.

22.

Mishra

Zhu

Corrections of Self-Selection Bias in Crash Causality Study: An Application on All-Red Signal Control. Journal of Transportation Safety & Security, Vol. 7, No. 2, 2015, pp. 107–123.

23.

Geedipally

S. R.

Patil

Lord

Examination of Methods to Estimate Crash Counts by Collision Type. Transportation Research Record: Journal of the Transportation Research Board, 2010. 2165: 12–20.

24.

Pendyala

R. M.

Washington

S. P.

Konduri

A Simultaneous Equations Model of Crash Frequency by Collision Type for Rural Intersections. Safety Science, Vol. 47, No. 3, 2009, pp. 443–452.

25.

Kim

D. G.

Washington

Modeling Crash Types: New Insights into the Effects of Covariates on Crashes at Rural Intersections. Journal of Transportation Engineering, Vol. 132, No. 4, 2006, pp. 282–292.

26.

Breiman

Friedman

Stone

C. J.

Olshen

R. A.

Classification and Regression Trees. CRC Press, Boca Raton, FL, 1984.

27.

Weng

Zhu

J. Z.

Yan

Liu

Investigation of Work Zone Crash Casualty Patterns Using Association Rules. Accident Analysis & Prevention, Vol. 92, 2016, pp. 43–52.

28.

AL-Zawaidah

F. H.

Jbara

Y. H.

Marwan

A. L.

An Improved Algorithm for Mining Association Rules in Large Databases. World of Computer Science and Information Technology Journal, Vol. 1, No. 7, 2011, pp. 311–316.

29.

Agrawal

Imieliński

Swami

Mining Association Rules between Sets of Items in Large Databases. Proc., 1993 ACM SIGMOD International Conference on Management of Data, Washington, D.C., 1993, pp. 207–216.

30.

Hornik

Grün

Hahsler

arules-A Computational Environment for Mining Association Rules and Frequent Item Sets. Journal of Statistical Software, Vol. 14, No. 15, 2005, pp. 1–25.

31.

Montella

Aria

D’Ambrosio

Mauriello

Data-Mining Techniques for Exploratory Analysis of Pedestrian Crashes. Transportation Research Record: Journal of the Transportation Research Board, 2011. 2237: 107–116.

32.

Abdelwahab

H. T.

Abdel-Aty

M. A.

Development of Artificial Neural Network Models to Predict Driver Injury Severity in Traffic Accidents at Signalized Intersections. Transportation Research Record: Journal of the Transportation Research Board, 2001. 1746: 6–13.

33.

De Oña

Mujalli

R. O.

Calvo

F. J.

Analysis of Traffic Accident Injury Severity on Spanish Rural Highways Using Bayesian Networks. Accident Analysis & Prevention, Vol. 43, No. 1, 2011, pp. 402–411.

34.

Besharati

M. M.

Tavakoli Kashani

Which Set of Factors Contribute to Increase the Likelihood of Pedestrian Fatality in Road Crashes?

International Journal of Injury Control and Safety Promotion, Vol. 25, No. 3, 2018, pp. 247–256.