Development Of An Objective Test And Evaluation System For Multi-object Recognition And Tracking Of Intelligent Vehicles

Vision-based multi-target detection and tracking is one of the research hotspots in the fields of computer vision and image information processing.It has important applications in driver assistance,driverless,and intelligent transportation.The development of multi-target detection and tracking algorithms,with the help of actual vehicle site testing,has high security risks,poor repeatability,and limited by road environment and experimental conditions,making it a multi-target detection and tracking algorithm for commercial operations or in the development stage.It has become an increasingly urgent need to carry out tests and objective evaluations,and to improve the algorithm through objective evaluation results.This paper takes the multi-objective detection and tracking algorithm based on machine vision as the research content,and designs the corresponding detection and tracking algorithm performance test evaluation system.The main contents are as follows:(1)The development of multi-target detection and tracking algorithms is based on real-vehicle site testing.There are problems such as high security risks,poor repeatability,and restrictions due to road environment and experimental conditions,such as hardwarein-the-loop and six-degree-of-freedom driving simulators.Based on this,a detection and evaluation system for intelligent vehicle environment-aware algorithms was developed.In order to take into consideration the repeatability and scalability of virtual testing and the authenticity of road tests,the real-time traffic scenes taken in advance were used as the input scene information of the system.(2)Aiming at the deficiency of the existing detection algorithm evaluation system based on manual calibration,a standard detection algorithm based on machine learning and a detection evaluation algorithm based on data association are proposed.The standard detection algorithm obtains the target detection and evaluation criteria by performing machine learning on the sample in the standard space in an off-line state,and then uses the test standard to test the detected image frame for comparison.The result association algorithm compares the detected results of each frame of the tested image with the standard detection results to determine the corresponding relationship between the test results of the algorithm to be tested and the standard detection results,and thus the accuracy of the multi-target detection results,the missing alarm rate,and other indicators The test evaluation is performed;the standard trajectory is generated by associating the data before and after the standard result,and the multi-target tracking trajectory of the image to be measured is measured and evaluated.(3)Using the method of subsampling to generate the standard trajectory through data correlation method,based on the data associating method to generate the standard trajectory,a trajectory matching and target matching test evaluation method based on ICP registration is proposed.The algorithm matching result is obtained the target trajectory.The distance error from the standard trajectory reflects the shape of the trajectory such as curvature,velocity,and trajectory,and the rotation and translation distance of each point on the curve.The tracking trajectories and standard trajectories input by the evaluation algorithm are represented by the main curve,and the main curve segment registration is realized by the ICP algorithm,so as to find the respective corresponding tracking trajectory segments of the multiple targets.According to the ICP algorithm results,the trajectory difference between the two can be evaluated.This value can be used as an evaluation index to evaluate the trajectory tracking performance.It solves the problem that the multi-target trajectory tracking is time consuming,the target is easy to lose,and the error rate of the target-dense area is high.(4)An index system for objectively evaluating the performance of the multi-target detection and tracking algorithm was established,and the multi-target detection and tracking test evaluation was performed through two aspects of the trajectory tracking evaluation index and the target detection and evaluation index.Among them,the tracking trajectory results are tested and evaluated in terms of accuracy,continuity and stability.The target detection and evaluation are evaluated in terms of accuracy,stability,versatility,and real-time performance.It solves the shortcomings of the current target detection tracking and evaluation indicators which are complex,the correspondence relationship is not clear,and only a single target is evaluated.