Research On Optimization Of Object Detection Algorithm In Highway Congestion Based On Deep Learning

In the field of computer vision,target detection is a very important branch.Its purpose is to extract deep features by using a large number of regular neural computing units for the input image or video,so as to get the categories and information of key targets in the image or video.This branch is a very important branch in computer vision.The target detection problem in the highway scene is a big problem.Due to the complex and changeable road conditions,shapes,vehicle density,and categories,the scene has the characteristics of perspective,extreme distribution of category size ratios,and serious occlu-sion problems.Use it directly Current target detection algorithms are often difficult to achieve a satisfactory result.Aiming at the problem of difficult detection caused by vehicle congestion in highway scenes,the target detection algorithm is researched and optimized.Combining the characteristics of occlusion and perspective,a new loss function and a new occlusion area are proposed respectively.The attention module optimizes the detection ability when vehicles are congested in highway scenes.The main work and related contributions of this article are sum-marized as follows:(1)According to the dependence of the extracted area effect and loss function in the can-didate area extraction structure in the target detection algorithm,for the occlusion problem between vehicles and vehicles in the highway congestion scenario,the occlusion information is combined with the loss function to give a loss function Weak supervision limitation of oc-clusion information,a target weight loss function based on target overlap is proposed.The results of training tests on congested data sets,non-congested data sets and public data sets show that after using the loss function proposed in this paper,we have achieved Compared with the current target detection algorithm,it has more accurate detection results and better subj ective effects,and the generated target frame and the official target are closer to each other,which has an improved effect on target detection in highway congestion scenarios.(2)According to the dependency relationship between the feature strength and the gen-erated result in the candidate region extraction structure in the target detection algorithm,for the occlusion problem between vehicles and vehicles in the highway congestion scenario,an occlusion area extraction module is proposed,which is in the neural network The obscured area is proposed,and the method of enhancing attention is used to extract the area information.The results of training tests on congested data sets,non-congested data sets,and public data sets show that after adding the module proposed in this paper,a result that is quite even better than the current target detection algorithm has been achieved.Target detection in the scene has a boosting effect.(3)Combined with the optimization method proposed in this paper,and using the key frame detection method to accelerate the calculation speed of the target detection algorithm,completed a real-time video detection vehicle detection algorithm in the highway congestion scene,completed the high-speed Event detection function under various scenes on highway.