Real-Time System for Tracking and Classification of Pedestrians and Bicycles

Data on pedestrian and bicycle volumes are necessary for transportation planning, infrastructure design, and traffic management. Nevertheless, such data cannot be collected directly by the commonly used detectors (e.g., inductive loops, sonar, and microwaves). In this study, a pedestrian and bicycle tracking and classification system was developed to detect pedestrians and bicycles with a video camera. This system contained six modules: a video flow capture module, a movement detection module, a shadow removal module, a feature extraction module, a tracking module, and a classification module. The Gaussian mixture model was used to extract moving objects from an image sequence. In the tracking module, the most challenging part of this system, the trajectories were obtained by use of a Kalman filter. To identify pedestrians and bicycles, a backpropagation neural network was used in the classification module. Two other simple but effective algorithms were used to alleviate the negative impacts of shadows and occlusions. The system was tested at three sites under different traffic and environmental conditions. It has been confirmed that the accuracy for pedestrian detection was approximately 85% and the count error rate was less than 13% for bicycles at all test sites. The proposed system is a feasible alternative for the collection of data for nonmotorized travel modes.