Research On Vehicle Small Object Detection Algorithm Based On Multi-Scale Vision

In recent years,Intelligent Connected Vehicle and Autonomous Driving technology have developed rapidly,and many universities and enterprises are constantly promoting the iteration and implementation of this technology.Environmental perception is an important technology for Intelligent Connected Vehicles to understand external information,and it is also a difficult part of autonomous driving technology.Vehicle detection based on visual information is an important part of environmental perception and one of the downstream tasks of object detection.However,the current vehicle detection technology still has the disadvantage of inaccurate detection of small objects.This is due to the large scale difference between objects,and the small objects of vehicles are dense and easy to form a mutual occlusion relationship.In response to this problem,this study takes the vehicle detection system based on monocular vision as the research object,and studies the multi-scale image detection algorithm,label assignment and Anchor Free method to improve the detection performance of the detection algorithm for small targets.The main research contents of this paper are as follows:(1)Design of vehicle detection algorithm based on multi-scale vision.This study analyzes the vehicle objects distribution in images of autonomous driving scenes,and proposes an image pyramid detection algorithm based on multi-scale vision to solve the problem of difficult detection of vehicle small objects set RoI.Based on the Local Outlier Factor algorithm,the corresponding ground truth labels of the small object set RoI are generated,which are used to train a lightweight detection network to specifically predict the small object area in the image.The training strategy is designed to enlarge the extracted area and detect in parallel with the original image.The output is fused to obtain an improved prediction result,which improves the detection performance of the model for small objects in the vehicle category.(2)Label assignment and Anchor Free method design based on FCOS model.After the mixed target set is enlarged and detected in the region of interest,the anchor box of the model is difficult to adapt to the enlarged region object when setting,and the label assignment strategy based on the anchor box limits the detection performance of small objects.In response to this problem,based on the FCOS model,the Anchor Free method is used to adaptively regress the prediction box expression,improve the SimOTA label assignment method,optimize the sampling of small objects,and improve the consistency of prediction classification and positioning.(3)Multi-Head structure design of the model output layer.Aiming at the low robustness of each grid cell adaptively predicting a single detection frame in the Anchor Free method,the Multi-Head structure of the prediction layer is designed in conjunction with Dropout.The prediction process can integrate Multi-Head without increasing the calculation amount of the prediction process.(4)Algorithm performance verification based on autonomous driving datasets and the vehicle collection data.This study uses the autonomous driving data set to train the algorithm,design experiments on the improved algorithm,verify the effectiveness of the algorithm,and verify the data collected based on the vehicle.The test results show that the improved vehicle detection algorithm in this study can significantly improve the detection of small vehicle objects and improve the overall detection performance.