Investigating the Influential Factors for Practical Application of Multi-Class Vehicle Detection for Images from Unmanned Aerial Vehicle using Deep Learning Models

Traffic density, which is a critical measure in traffic operations, should be collected precisely at various locations and times to reflect site-specific spatiotemporal characteristics. For detailed analysis, heavy vehicles have to be separated from ordinary vehicles, since heavy vehicles have a significant effect on traffic flow as well as traffic safety. With unmanned aerial vehicles (UAVs), it is easy to acquire video for vehicle detection by collecting images from above the traffic without any disturbances. Despite previous studies on vehicle detection, there is still a lack of research on real-world applications in estimating traffic density. This study investigates the effects of several influential factors: the size of objects, the number of samples, and a combination of datasets, on detecting multi-class vehicles using deep learning models in various UAV images. Three detection models are compared: faster region-based convolutional neural networks (faster R-CNN), region-based fully convolutional network (R-FCN), and single-shot detector (SSD), to suggest guidelines for model selection. The results provided several findings: (i) vehicle detection from UAV images showed sufficient performance with a small number of samples and small objects; (ii) deep learning-based multi-class vehicle detectors can have advantages compared with single-class detectors; (iii) among all the models, SSD showed the best performance because of its algorithmic structure; (iv) simply combining datasets in different environments cannot guarantee performance improvement. Based on these findings, practical guidelines are offered for estimating multi-class traffic density using UAV.