基于网络流跟踪的信号灯检测方法

Detection of traffic lights for off-line videos based on network flow tracking

结合信号灯信息对机动车行进速度进行引导,减少机动车启停次数,可有效减少废气排放,缓解其造成的污染问题。针对信号灯转换时刻的获取问题,提出了一种基于网络流跟踪的信号灯检测方法。首先,该方法在数据集中引入辅助信号灯类别进行训练,将视频序列中该类目标检测结果关联为踪片,并通过踪片建模多目标跟踪任务。其次,该方法将多目标跟踪任务转换为最小费用流优化任务,以踪片作为节点建立最小费用流网络,提出了适合于信号灯的费用构建方式,通过最短路径算法求解,得到视频序列中辅助信号灯的多条轨迹。最后,基于求解的轨迹结果和图像分类技术,实现信号灯检测性能的提升。该方法的跟踪性能相较于对比算法有大幅提升,并将小目标信号灯检测响应的mAP提升至94.35%。实验结果表明,基于网络流的建模方式能极大地提升信号灯的跟踪准确率,结合跟踪轨迹还能大幅提高视频序列中小目标信号灯的检测准确率,并可有效确定信号灯状态的转换时刻。

Combining the traffic light information to guide the speed of vehicles and reducing the number of starts and stops of vehicles can effectively reduce exhaust emissions and alleviate pollution problems caused by them.Aiming at the acquisition of traffic light transition time,this paper proposed a traffic light detection method based on network flow tracking.It introduced auxiliary traffic light class in the dataset for training,and correlated the detection results of this class in the video sequence as a tracklet,then modeled the multi-object tracking task by the tracklet.Specifically,this method converted the multi-target tracking task into the minimum cost flow optimization task.This method established the minimum cost flow network with the tracklet as the node,and proposed a cost construction method suitable for traffic lights,then obtained multiple traces of the auxiliary traffic light in the video sequence by solving the shortest path algorithm.Based on the solved trajectory results and image classification technology,it finally improved the performance of traffic light detection.Compared with the comparison algorithms,the proposed method greatly improves the tracking performance,and increases the mAP of the small target traffic light detection response to 94.35%.Experimental results show that the network flow modeling method can greatly improve the tracking effect of the auxiliary traffic light.Combined with the tracking results,it greatly improves the detection accuracy of the small target traffic light in the video sequence and effectively determines the transition time of traffic light status.