面向车辆自主驾驶的行人跟踪算法

Pedestrian tracking algorithm for autonomous driving

基于背景感知相关滤波框架和车辆前方行人运动的特点,建立了运动行人尺度快速估计和选择性更新的行人跟踪算法。首先,在线训练学习待跟踪行人的背景感知相关滤波器。其次,针对行人的尺度变化训练一个一维的尺度相关滤波器对尺度进行精细搜索,使算法更适应车载的快速尺度变化。再次,利用峰值旁瓣比评价行人状态,建立两相关滤波器的选择性更新机制。最后,基于吉林大学车载试验数据库JLU-PDS、德国奔驰Daimler、美国OTB共享国际测试库,与卡尔曼车载行人跟踪算法进行对比测试,试验结果表明本文算法具有较好的尺度适应和抗遮挡性能,更能满足车辆自主驾驶的需求。

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.