公路视频实时车辆检测分类与车流量统计算法

Highway video real-time vehicle detection classification and traffic flow statistics algorithm

公路视频实时车辆检测分类与车流量统计是计算机视觉领域的一个经典问题。传统设置检测带法,易漏检复检,自动化性不好。基于深度网络的one-stage算法实时性好,但是经常会把变化的背景、运动的非车辆物体纳入其中,同时对光照变化敏感,夜间分类效果不好。因此,提出采用one-stage做目标检测,并不直接获取分类结果,而是根据标注框将物体切割出来,去除背景,提升抗背景扰动性能和分类效果;再送入一个经过迁移学习的浅层神经网络;将分类输出和目标检测网络的位置输出合并送入一个全图匹配算法,进行车流量统计。该方法在保障实时性的同时降低了漏检和复检率。

Real-time vehicle detection,classification and traffic statistics based on road video are classic problems in the field of computer vision.The traditional method of setting the detection belt is prone to missed inspection and re-inspection,so the automation performance is not good.The real-time performance of one-stage algorithm based on deep network can be guaranteed,but the changing background,moving non-vehicle objects are often included,and the change of illumination is sensitive at the same time,so the classification at night is not good.Therefore,an algorithm is proposed to perform target detection by one-stage,and the classification result is not directly obtained.Instead,it cuts out the object according to the bounding box,removes the background,and improves resistance to background disturbances and classification accuracy.Then it is sent to a transfer learning shallow neural network.The classified output and the position output of the target detection network are combined and sent to a full map matching algorithm for traffic flow statistics.While ensuring real-time performance,the rate of missed inspections and re-inspections is reduced.