Research On Video Object Detection In Complex Traffic Scenes

As a popular engineering discipline,computer vision has been more widely used in industrial production and daily life with the rise of convolutional neural networks.Among them,as one of the important branches of computer vision,video object detection has been greatly developed in recent years,and it also plays an extremely important role in the field of intelligent transportation.At present,most object detection algorithms can detect objects well in the normal environment,but the robustness in complex scenes or extreme conditions still need to be improved.In addition,as for some distant objects which are usually mixed with the background,the response of these objects on the feature map cannot be distinguished,which greatly limits the performance of conventional detection algorithms.At the same time,since most of the video object detection algorithms are too complicated too achieve real-time detection,and it is difficult to apply it to real life.Therefore,achieving video object detection in complex scenes is a hot and difficult point.In view of the above problems,this paper starts from two aspects of detection accuracy and speed:(1)Aiming at the problem that the far-end object in the video frame of the complex traffic scene is easy to merge with the background,resulting in poor detection accuracy,we propose a new video object detection framework.The framework combines traditional moving object detection methods with mainstream convolutional network detection algorithms,and consists of an appearance stream and a movement stream,which are used to mine the appearance information and movement information of the object,respectively.At the same time,in order to solve the problem that the object is lost in subsequent frames due to occlusion or motion blur,we introduce a memory attention module to capture the temporal correlation in adjacent video frames.In dual-stream networks,we use mainstream detection frameworks to extract appearance information,and motion information is extracted using the frame difference method in traditional moving object detection.After combining the appearance flow with the movement flow,the appearance flow makes up for the defect that the movement flow cannot detect the stationary object,and the movement flow improves the response of the movement object in the feature map.In the memory attention module,we introduce the attention mechanism into the cyclic convolutional network,and use the state of adjacent frames to significantly improve the detection effect of occlusion and blurred objects.(2)For the problem of slow inference speed of convolutional neural networks,we first proposed a knowledge distillation object detection algorithm NSNet(Narrow ShuffleNet v2)based on deep separable convolutions.NSNet is a lightweight network that includes a point-wise group convolution and a bottleneck structure.By shrinking and expanding on the two components of the dimension and space scale of the feature map,the calculation speed and accuracy are balanced.Finally,a knowledge distillation method is designed to improve the detection accuracy while keeping the detection speed constant.