Research On Road Target Detection And Location For Vehicle Forward Collision Warning

In recent years,with the continuous development of deep learning technology and binocular vision technology,it has shown excellent performance in the field of object detection and localization.In the field of intelligent vehicles,the detection and localization of road targets in traffic scenarios is an important component of the vehicle intelligent driving assistance perception system.Currently,the cost of object detection and localization schemes based on Li DAR and millimeter wave radar is relatively high,making pure vision based object detection and localization schemes a research hotspot.Therefore,this article uses binocular cameras to obtain road target information such as pedestrians,motor vehicles,and non motor vehicles in front of the vehicle,uses deep learning based object detection algorithms to obtain target category information,and combines binocular vision technology to obtain target position information.Based on this,a vehicle front collision warning system is studied.This paper mainly relies on the national key research and development program "Research and Practice of swarm intelligence Checked Vehicles for Airport Baggage Transfer"(2021YFE0203600).The main research contents are as follows:(1)Analyzed the principle and structure of YOLOv5s object detection algorithm,providing a basis for subsequent improvement strategies of YOLOv5s object detection algorithm;The advantages and disadvantages of local stereo matching methods,global stereo matching methods,and semi global stereo matching methods were compared through experiments;To address the issue of noise generation in the SGBM(Semi-Global Block Matching)algorithm,a WLS(Weighted Least Squares)filter is used to improve and optimize it.(2)In the road object detection section,in response to the problem of poor detection accuracy in YOLOv5s network,the CA attention concentration mechanism was added to the network backbone.Comparative experiments showed that the improved algorithm effectively improved the network’s recognition rate for small and occluded road targets;In response to the problem of poor regression accuracy of YOLOv5s target box,we introduced the α-CIoU loss function calculates the position of the prior frame of the target.Comparative experiments show that the introduction of α-CIoU loss function effectively improves the regression accuracy of the target detection frame.The ablation experiment has proven that YOLOv5s network incorporates CA attention mechanism and α-CIoU loss calculation function,the precision of the target was improved by 0.9%,the recall rate was improved by 2%,and the mean average precision was improved by 1.1%.(3)In the target localization section,the SGBM algorithm with balanced performance is selected as the basic algorithm for road target localization.Based on the improved SGBM,a road target position coordinate calculation architecture is designed.The experiment shows that the maximum absolute error of target location is 0.46 m and the maximum relative error is 2.2%within the effective detection range of 22 m of the binocular camera.(4)Considering the joint task of target detection and target positioning,the fusion of target detection algorithm and target positioning algorithm has achieved the output of road target category,3D coordinates,and distance information.Finally,a vehicle front collision warning model based on TTC(Time to Collision)was designed based on the fusion network,and experimental verification was conducted.