Vehicle Detection Methods And Their Applications In Open Scenarios

Object detection has always been a hot topic in the field of computer vision,and deep neural networks have been effective in traditional object detection tasks.However,in practical scenarios,object detection faces many challenges,such as different detection angles,obscured objects to be detected,and limited computing power of mobile devices,which result in poor detection performance of the algorithm model.But as object detection technology continues to improve,more and more technologies are being applied to object detection algorithms to meet people’s increasing demand for detection efficiency and accuracy.The main research direction of this article is to apply deep learning methods to detect various indicators of vehicles,including vehicle color detection and vehicle position localization.Currently,this research topic mainly faces two major challenges: one is how to achieve accurate vehicle detection in complex practical environments,and the other is how to achieve rapid vehicle detection on low computing power devices.To solve the above difficulties,this article studied relevant algorithm models and implemented an algorithm system,and extended the algorithm model in the experimental environment to real scenarios.The main contributions of this article include the following three points:(1)A lightweight vehicle color detection method based on YOLOv5 s is proposed.Three lightweight networks are combined with the YOLOv5 s model to solve the problem of low device computational power.Specifically,three lightweight networks are combined with the feature extraction part of YOLOv5 s to further reduce the model parameter size and improve the model detection speed without reducing detection accuracy.Mobile Netv3 uses depthwise separable convolution to greatly reduce the model’s computational complexity;Shuffle Netv2 uses channel shuffling to perform channel reordering and achieve cross-group information exchange;Ghost Net attempts to obtain redundant feature maps using lower cost computational resources.These three networks achieve lightweight models from different angles.Experimental results show that compared to the original YOLOv5 s model,the algorithm model with lightweight networks achieves a faster detection speed of 125 fps and a minimum model parameter quantity of one-fifth of the original network.The accuracies of the three lightweight models are 96%,94.4%,and 95.8%,respectively.By adding a Bi FPN structure to the feature fusion part of the YOLOv5 s model,multi-scale feature fusion is achieved to further improve detection accuracy and speed.Based on the lightweight network,the accuracies of the three lightweight networks are 96.1%,95.9%,and 96.1%,respectively,after adding the Bi FPN structure.(2)A vehicle color detection method that combines attention and self-attention mechanisms is proposed.Various attention and self-attention mechanisms are added to the YOLOv5 s model,and ACON activation function and SIo U loss function are introduced.Specifically,CBAM,SE,and CA attention mechanisms and Bo TNet and Co TNet self-attention mechanisms are added to the original network,and ablation experiments are conducted to compare the performance improvement of various attention and self-attention mechanisms.(3)A vehicle recognition and detection system based on deep learning methods is developed,and the improved algorithm models are applied to practical scenarios.The system includes a Web interaction module,a vehicle appearance detection module,and a vehicle location positioning module,where two different models can be selected to detect input data in the vehicle appearance detection and location positioning modules.Finally,through system testing,the completeness and compatibility of the system’s functions are demonstrated.In summary,this thesis proposes two solutions to the problems of vehicle detection: one is based on lightweight networks and bidirectional weighted feature networks,and the other is based on attention and self-attention.