YOLOX-Based Binocular Vision Forward Vehicle Tracking And Ranging Research

At present,the total number of cars worldwide continues to increase,which on the one hand facilitates people’s daily travel needs,but on the other hand brings about very many traffic safety problems.Due to the high speed of cars on the road,drivers are unable to react in time to some emergencies,which leads to traffic accidents.Ongoing studies on Driverless technology and ADAS aim to support drivers in assessing road conditions and making emergency decisions.This approach reduces reliance on drivers,leading to improved safety and decreased incidence of traffic accidents.Real-time detection and tracking of vehicles in front of you as well as distance measurement and positioning technology is the core of the development of autonomous driving technology.Many current vehicle detection algorithms can do a good job of detecting vehicles in simple detection scenarios,but if there are occlusions,targets that are too small,weather changes and other situations,the detection effect will be greatly reduced or even invalidated.Moreover,most vehicle detection algorithms are poor in real time,and problems such as missed detections and jittering detection frames can occur during the detection of video streams,which cannot be consistently applied in real road scenarios.In addition,most of the current distance measurement sensors are expensive.In response to the above problems,author put forwarded a YOLOX-based dual present vehicle tracking and ranging method,using the improved YOLOX model for vehicle target detection,using the Deep Sort algorithm for vehicle target tracking,and then using the binocular camera to range the vehicle ahead.In addition,a binocular vision system is built,and the advantages of this method are demonstrated and analyzed through timely vehicle ranging and tracking experiments.The specific research contents are as follows:(1)To tackle the issues of low detection accuracy and inadequate real-time performance in many existing vehicle detection algorithms,this paper enhances the YOLOX algorithm to tackle these problems.To reduce the size of the model and improve the real-time performance of the model,the Ghost module was used to optimize the YOLOX backbone feature extraction network for lightness.To solve the problems of vehicle occlusion and different scales of vehicle targets under complex roads,the feature fusion network of the model is optimised by extending the three effective feature layers of the input feature fusion network to four scales,increasing the 2-fold upsampling fusion to 4-fold,and performing jump connections.The modified model was validated using a self-made vehicle target dataset and it can be obtained that the optimized YOLOX model has obtained 90.2% accuracy value on the self-made vehicle data set,1.6percentage points higher than the original YOLOX,and the detection time for a single image is22.8ms,22.4ms lower than the original YOLOX.(2)The Deep Sort-based vehicle tracking algorithm is proposed to address the instability of vehicle detection algorithms when performing real-time vehicle detection.Train my own vehicle re-identification model using the Ve Ri-776 vehicle re-identification dataset,Combining Deep Sort with an improved YOLOX and validating the effect on the KITTI tracking dataset.it can be obtained that a single target recognition algorithm is unstable when detecting the target vehicle in the video frame,and there will be a lot of missed detection.Combining the tracking algorithm can improve the stability of the detection effect and effectively improve the missed detection.The algorithm obtained 90.6% of MOTA and 91.3% of MOTP.(3)Aiming at the problem that the traditional ORB feature matching algorithm has many mismatches,an ORB feature matching algorithm based on PROSAC optimization is put forwarded.Numerous false matches occur when using only ORB for feature matching.Therefore,to solve this problem,In this paper,PROSAC is used to improve this situation and reduce the number of feature points incorrectly matched by ORB algorithm.It was also verified experimentally and the results showed that the improved ORB algorithm has a ranging error of3% to 5% in the range of 20 to 34 meters,which is 1% to 2% lower than before the improvement.In this paper,a binocular vision system is built and experiments are carried out for realtime forward vehicle tracking and forward vehicle distance measurement respectively.The results show that the algorithm can control the average error rate of the forward vehicle distance measurement within 3% in the distance range of 0 to 10 m.Good results can also be achieved in real-time vehicle tracking tasks.It is verified that the method proposed in this paper can satisfy the detection,tracking and ranging of forward vehicles under normal road scenarios.