Research On Complex Scene Vehicle Detection Algorithm Based On Improved YOLO

With the development of China’s economy,the vehicle ownership rate is increasing year by year,leading to increased traffic density,which in turn leads to road congestion and frequent accidents.How to meet the high traffic management demand with limited human resources is an urgent problem.To solve this problem,many scholars have developed solutions such as smart transportation and driverless based on the development of deep learning and computer hardware.Vehicle detection,an important detection aspect of smart transportation and autonomous driving,is a hot research topic.Vehicle detection relies on target detection models to accurately identify the location and class information of vehicles in different pictures.From two-stage to single-stage target detection models,the speed and accuracy are continuously balanced,but they are still inadequate for some specific problems.In this paper,we summarize and explore the current stage of vehicle detection models,and make improvements and innovations.The main work is as follows.(1)In order to reduce the leakage and false detection of targets in the detection process and to better detect objects of different scales in images,this paper proposes a vehicle detection algorithm based on the YOLOX model.Firstly,the data set is processed using the Mosaic+Mixup data enhancement method,which increases the number of small targets,enriches the background information and improves the robustness of the model;next,the ECA attention mechanism is added to the backbone network to reduce the impact of complex background information on the detection effect;finally,the Bi FPN is used as a feature fusion network to improve the model’s ability to fuse advanced features The final use of Bi FPN as a feature fusion network improves the ability of the model to fuse advanced features.After extensive experiments on the UA-DETRAC dataset and the KITTI dataset,the results show that the detection performance of the MEB-YOLO model is better than the existing cash methods.(2)In order to enable the model to better detect vehicles on more complex scenes such as factories,rural areas and construction sites in real time,we collected more than30,000 images from the above scenes and combined them with the UA-DETRAC dataset into a new dataset.In this paper,we propose a new vehicle detection algorithm SMAYOLO based on the YOLOv7 model.First,we use Mobile Net V3 as a new feature extraction network to lighten the model,next,we use Sim AM attention mechanism to reduce the effect of background information on the detection effect,in addition,we use ACON activation function to replace the original Si LU activation function in YOLOv7 model,and finally,we use CIo U to replace SIo U to optimize the loss function.This model achieves the effect of model lightweighting,significantly reducing the size,computation and number of parameters of the model,while improving the detection speed and accuracy of the model,providing the possibility of deploying the detection model to embedded devices for real-time detection.