Research And Application Of Vehicle Detection Algorithm Based On Deep Learning

Vehicle detection is one of the research topics in target detection and an important foundation for digital image processing tasks.Deep learning-based detection algorithms are still good for detecting targets with uniform distribution,but the detection accuracy is not high in the case of dense vehicles or vehicles with occlusion,and the detection requires large computing power.The complex network layer of the deep learning-based target detection algorithm requires strong computational power,which causes a decrease in the detection efficiency of the algorithm and difficulty in the deployment of the algorithm.The existing detection algorithm is now leveraged as the basis to study the corresponding detection network in terms of reducing algorithm miss detection and reducing the number of algorithm parameters,mainly as follows:(1)A proposal for a vehicle detection algorithm featuring cross-attention and feature scale fusion is put forward to address the issue of low detection accuracy,caused by the abundance of overlapping information of dense targets and sparse edge transitions.First,VGNet G network is employed as the backbone replacement to decrease the computational parameters of the model by VGNet G network;additionally,FSSA module was implemented to further join the backbone information features,thus allowing the model to be more comprehensive.contextual information is aligned with the up-sampled high-level feature information;finally,the improved cross-attention module ICCAttention(Improved Criss Cross Attention)utilizes two connection maps instead of the common single dense connection map,and efficiently extracts the contextual information of all pixels by the cross-path method.The YOLOv3 model’s detection of dense targets is significantly enhanced by the improved algorithm,and it is able to meet real-time detection speed requirements without compromising accuracy.(2)A Half-bottleneck Convolutions(HC)module,improved to replace the C5 module based on YOLOv7 Algorithm,is proposed to address the issue of missed detection caused by scale change and the large number of parameters that are not easily deployed.This asymmetric attention and separation detection algorithm is adoped to enhance the feature extraction capability of the backbone network,which can obtain a rich stream of feature gradients with high efficiency;second,the attention module 3D Irregular Non-local Attention Module(3D INAM)is utilized to expand the model perceptual field and ignore the background information,dig deeper into the detection object feature information,increase the multi-scale features and enhance the model feature extraction capability.The improved algorithm achieves 82.9% accuracy on the new data set,which effectively improves the leakage detection problem of the original algorithm.(3)Validation of the enhanced algorithm on the UA-DETRAC target dataset,obtained at intervals,has been conducted,and the experimental results demonstrate that the vehicle detection algorithm with cross-attention and feature scale fusion is successful in detecting dense targets and meets real-time performance.Additionally,the asymmetric attention and separated target algorithm has been found to significantly enhance detection accuracy with a certain degree of robustness.With the software deployment algorithm to the mobile terminal,the code of vehicle detection and meow reminder function are combined,when the detection is completed by turning on the camera or capturing the photo directly,the detection result will be pushed directly to the cell phone with the quantity information,which can understand the expected traffic flow of the forward route,and then provide the driver with pre-selected routes for the driver to better plan the forward route.