Abstract:

Pedestrian tracking is the base for pedestrian behavior analysis and driving decision?making. There exist several problems such as the fast scale change of pedestrian caused by rapid movement of vehicle, the pedestrian occlusion, etc, which make the traditional tracking algorithm difficult to track pedestrians accurately, and it is more difficult to analyze pedestrian movement. To overcome the problems, a new pedestrian tracking method is presented in the paper. In the method, scale estimation and selective updating strategy were used to deal with fast scale change of pedestrian and occluded in the framework of background?aware correlation filter. First, a background?aware correlation filter was trained online for the pedestrian to be tracked. Secondl an one?dimensional scale correlation filter was trained to search the scale of pedestrians carefully, so that the algorithm is more adaptable to fast scale change in the case of vehicle driving. Finally, a selective updating mechanism of the correlation filters was set up by using the peak sidelobe ratio to evaluate the pedestrian status. In experiments, the proposed method was compared with Kalman filter pedestrian tracking algorithm on JLU?PDS, Daimler Pedestrian Benchmark Data Set and some video sequences from Object Tracking Benchmark. The results show that the proposed method has better scale adaptation and anti occlusion performance, and it is more adaptable to the autonomous driving.

Key words: engineering of communication and transportation system, autonomous car, pedestrian tracking, background?aware correlation filter, scale estimation, selective model updating