Research On Vehicle Object Detection Method Based On Deep Learning

In recent years,with the surge in the number of motor vehicles in our country,the pressure of government supervision is increasing,and traffic accidents occur frequently.Many researchers have tried to solve the problems caused by the proliferation of vehicles by applying artificial intelligence technology to vehicle driver assistance systems and traffic supervision systems.As a key technology for the implementation of vehicle driver assistance systems and traffic supervision systems,vehicle target detection technology should have high detection accuracy and be suitable for complex and diverse detection scenarios.In addition,with the continuous development of Internet of Things technology,it has become a trend to deploy vehicle target detection algorithms on mobile devices.Due to the small memory and insufficient computing resources of mobile devices,vehicle target detection algorithms need to be lightweight to meet the deployment conditions of mobile devices.This paper constructs a vehicle target detection method based on YOLOv5,and optimizes it from two aspects of improving detection accuracy and achieving model lightweight.The main research work is as follows:(1)The accuracy of the vehicle target detection method based on YOLOv5 is improved.The K-means ++ algorithm was used to optimize the initial anchor box to improve the positioning ability of the model for the vehicle target.The SIoU loss function was used to calculate the positioning loss to improve the positioning accuracy of the model.The SimAM non-parametric attention mechanism is integrated into the backbone network to improve the vehicle target detection accuracy without increasing the amount of additional parameters.(2)In order to improve the detection performance of the model for small target vehicles and occluded vehicles,optimization research was carried out to solve the difficult problems in the detection of two types of vehicle targets: small target vehicles and occluded vehicles.In the detection of small target vehicles,a new feature extraction network is constructed in the backbone network by using the idea of dense connection to enhance the feature extraction ability of the model,and a weighted bidirectional pyramid network structure is used for multi-scale information fusion in the neck network to further enhance the representation of vehicle target features.In the detection of occluded vehicles,the Non-Maximum Suppression algorithm is improved to the Soft-SIoU-NMS algorithm to reduce the false detection and missed detection of vehicles under dense and occluded conditions.(3)The vehicle target detection model is lightweight improved.GhostNetV2 network was used to improve the backbone network,and two improvement schemes were proposed to build a new feature extraction module C3＿GhostV2 and overall replacement,respectively.The feature extraction module C3＿GhostV2 with better performance was selected through experiments to realize the lightweight improvement of the backbone network.The ghost shuffling convolution module is used to replace the ordinary convolution module in the neck network to further reduce the number of parameters and improve the operation speed of the vehicle detection model.