Research On Traffic Target Detection Algorithm Based On Deep Learning

With the proliferation of car ownership,road congestion,safety accidents and other traffic problems are becoming increasingly prominent,and intelligent transportation,assisted driving and autonomous driving technologies,as important technologies for solving traffic problems,are receiving more and more attention and research from the state and enterprises.Target detection technology,as one of the key technologies,will directly affect the intelligence level of the subsequent decisionmaking system and control system.In the target detection task of actual traffic scenes,there are problems of large target scale gap,many dense and small targets,serious intertarget occlusion and complex background interference,which lead to serious leakage and insufficient detection accuracy of conventional target detection algorithms.Therefore,this thesis proposes a multi-scale feature-enhanced traffic target detection algorithm incorporating attention mechanism with YOLOv4 algorithm as the basic framework and the difficulties and requirements of target detection in real traffic scenes to achieve high detection accuracy while ensuring real-time detection,as follows:(1)To address the problems of large target scale disparity,dense small targets and mutual occlusion between targets in actual traffic scenes,this thesis proposes a multiscale feature-enhanced traffic target detection algorithm using YOLOv4 as the baseline algorithm for improvement.Firstly,in order to enhance the multi-scale feature extraction capability of the backbone feature extraction network,a multi-scale feature extraction module is constructed to enhance the extraction capability of the backbone feature extraction network for different perceptual field features through three cavity convolution branches with different expansion rates;considering the problem that the SPP module is prone to feature loss during the downsampling process,the RFB module is introduced to replace the SPP to obtain the global perceptual field in the deep layer of the network.In order to enhance the network’s ability to detect small targets,the network is also able to extract features from different receivers.Then,in order to enhance the detection capability of small targets,the feature fusion network is expanded to four layers by adding an extra small target detection layer for detecting extra small targets;at the same time,since the multi-layer convolutional stacking of the feature fusion network will lose the original feature information,two cross-layer connections with adaptive learning weights are added to connect the backbone network feature information to the PAN part of the feature fusion network to better fuse the feature information.Finally,in order to solve the problems of occlusion target miss detection and sample imbalance,flexible non-extreme suppression and focal loss are introduced for optimization.The multi-scale feature-enhanced traffic target detection algorithm is validated on the KITTI dataset and the UA-DETRAC dataset respectively,and the experimental results show that the improved algorithm in this chapter is more effective in detecting dense occluded targets.(2)To address the problems of complex backgrounds and serious interference in real traffic scenes,the multi-scale feature-enhanced traffic target detection algorithm is further optimised on the basis of the multi-scale feature-enhanced traffic target detection algorithm,combined with the attention mechanism,to construct a multi-scale feature-enhanced traffic target detection network incorporating the attention mechanism.Firstly,the existing attention mechanism is studied,and for the large number of parameters and serious loss of features in the process of dimensionality reduction in the fully connected layer of CBAM,while the channel attention mechanism and spatial attention mechanism are connected in tandem to easily cause mutual interference between features,a parallel and efficient attention module ECBAM is proposed,which has better performance on the premise of saving the number of parameters and speeding up the operation;then the self-attentive mechanism is The ACB module is then combined with ACmix to construct an ACB module,and the ACB module is added to the backbone feature extraction network to construct inter-target connections through inter-pixel dependencies,which can be combined with the localization of convolution to more fully exploit the effective information of the feature map;finally,the improved attention module and the improved self-attention module are introduced into the multi-scale feature-enhanced traffic target detection network,and the depth-separable convolution is also introduced to further Finally,the improved attention module and the improved self-attentive module are introduced into the multiscale feature-enhanced traffic target detection network.The experimental results show that the final optimized traffic target detection algorithm has better detection performance.The thesis has 61 figures,9 tables,and 86 references.