Research On Traffic Target Detection Methods In Road Scene

As the number of vehicles on the road increases,traffic problems become more severe.Therefore,deep learning-based object detection algorithms are increasingly being used in the traffic field.Vision-based object detection algorithms have developed rapidly and are comparable to or even exceed human-level performance for general road scenes.However,in complex traffic scenes with densely distributed objects and occlusion,there is a high chance of object omission,which makes it difficult to achieve the required detection accuracy.To address this issue,this thesis proposes two object detection algorithms based on deep learning.(1)Aiming at the problem that it is difficult for general networks to extract effective features in complex traffic scenes,this thesis proposes a new object detection network model,Retransformer-YOLOv5,based on YOLOv5 X,with the efficient feature extraction network as the starting point.The network first introduces swin-transform with stronger feature extraction ability as a new backbone network,and then improves its self-attention mechanism to solve the attention collapse phenomenon in the deep attention map;Then,in the post-processing stage of the network,a post-processing method of weighted frame fusion is introduced.By weighted fusion of all prediction frames with high confidence,the problem that the network is difficult to output the correct prediction frame due to mutual occlusion of targets in complex traffic scenes is solved,making the network more suitable for traffic object detection in complex traffic scenes.(2)Aiming at the problem that the lightweight object detection algorithm is difficult to take into account both speed and accuracy in traffic scenes,this thesis proposes an efficient lightweight object detection algorithm LODN in Chapter 4.LODN first uses Depth-wise separable convolution to build a lightweight feature extraction network.At the same time,in order to ensure the feature extraction ability of the lightweight backbone network,an improved coordinate attention mechanism is embedded in it,so that the network can more accurately locate the location information of the object of interest,thus improving the recognition ability of the model;Secondly,an adaptive scale feature weighted fusion strategy is proposed to measure the importance of different feature layers in the path aggregation network feature pyramid,so that the network can focus on learning the effective information of different feature layers in the pyramid.Compared with YOLOv4,LODN achieves more balanced results between model size and accuracy.Experiments on the BDD-100 K dataset show that LODN achieves 87% of YOLOv4 in m AP while controlling the parameters and calculation complexity of the model to within 15% of YOLOv4.