Research On Vehicle Visual Field Object Detection Method Based On Deep Learning

In recent years,vehicle visual field target detection has become a hot issue in intelligent driving systems.Based on the advanced one-stage target detection method,it is a research trend in computer vision to give vehicles the ability to autonomously perceive the road and the surrounding environment,and to accurately identify and respond to the target ahead.Therefore,as the core basis of unmanned driving technology,vehicle visual field target detection method has important research significance.1.A local fragile field enhancement model based on YOLOv3 is established.Using the Euclidean distance as the metric,the improved K-median clustering method is used to obtain a priori anchor box with higher matching degree;Aiming at the problem that small target features are sparse and difficult to extract under long-distance vision in images,a local fragile region enhancement method is proposed to sort out and enhance local chaotic features,and use parallel branches to avoid the loss caused by deep convolution down-sampling;In order to reduce the insufficiency of the YOLO series center point prediction rules,the output end is enhanced and corrected by flexible sampling of surrounding grid features and weighted analysis to improve the reliability of the coding features corresponding to the center grid,and a grid offset sampling strategy is proposed.Finally,after a large number of ablation experiments and comparative analysis with other methods,it is proved that the model has higher detection accuracy.2.Aiming at the problem of uneven distribution of multiple scales and the difficulty of distinguishing multiple congested targets at the same time by the central grid feature tensor,an anti-congestion vehicle detection model based on YOLOv4 is designed to improve the recognition accuracy of the network for dense long-distance small targets.First,use the improved K-median clustering algorithm to get more accurate a priori anchor boxes.The foreground and necessary background features are extracted through the proposed context module to enhance the model’s ability to perceive the surrounding environment of the target;At the same time,in order to make the model obtain powerful multi-level features,this thesis proposes a prediction layer that focuses on balance;finally,in order to achieve accurate distinction of congestion targets,an anti-congestion module that relies on strict evaluation criteria is designed,combining parameters such as quantity and distance.To define the evaluation score,this module performs secondary detection on local areas that are easily missed.The model is compared with a variety of prediction methods,and the results show that the method has better performance on dense detection problems.