基于视频时空关系的高速公路异常停车检测

Highway abnormal parking detection based on video spatio-temporal relationship

针对传统高速公路异常事件检测方法效率低、漏检率高、实时性较差等问题,提出基于视频时空关系的高速公路异常停车检测方法。首先,采用基于交通流频率分析的无监督分割方法分割道路图像,去除较小连通域提取道路分割图;然后,通过透视关系模型将近远目标归一化到同一尺度并进行裁剪,输入YOLOv4网络进行二次检测增强对近远目标检测的鲁棒性;最后建立时空信息矩阵,通过时空矩阵的更新与NMS(non-maximum suppression)方法检测合并异常区域并输出检测结果。实验结果显示,该方法在远距离场景中的准确率为95%,在拥挤场景中的准确率为93%;通过对比实验结果发现该方法能够有效提高复杂场景下异常停车的检测准确率且具有良好的泛化能力。

The current highway abnormal event detection efficiency is low,the leakage rate is high,real-time poor,and so on,this paper proposed highway abnormal parking detection method based on space-time relationship.Firstly,it used the traffic flow frequency to quantify the split image and removed the smaller connectivity domain to extract the road surface information,then used the perspective geometry of the pinhole camera to normalize the same scale.It used YOLOv4 network to obtain the exact position of the vehicle target,and established the space-time information matrix.To detect abnormal region and the detection results can output by the update of the space-time matrix.The experimental results show that the accuracy of this method is 95%in long-distance scenes and 93%in crowded scenes.