Research On Extended Target Tracking Technology Of Millimeter Wave Radar In Road Scene

In recent years,with the increasingly complexity of the road,the role of Intelligent Transport System（ITS）has become increasingly prominent.Millimeter-wave radar has become one of the core sensors in ITS due to its advantages of all-day,all-weather,and high-resolution.Designing a multi-target tracking algorithm with the highest possible accuracy and the best real-time performance based on the complex point Clouds in the actual application environment is of great significance for improving the performance of ITS.To this end,the thesis focuses on the extended target tracking technology based on millimeter-wave radar in road scene and verified the performance by the measured point Clouds.The specific research are as follows:（1）Aiming at the problem of clustering in the process of extended target tracking,the thesis proposed the DBSCAN algorithm based on target motion characteristics suitable for the near field.The algorithm determines whether to cluster again by judging whether the initial clustering result of DBSCAN is a double-target point cluster,which effectively solves the problem of the DBSCAN algorithm in multi-target scene,which is caused to cluster inaccurately by the short distance and the similar motion states of the double targets.At last,the thesis verified the clustering performance of the algorithm by the measured point Clouds in the near field.（2）In the process of extended target tracking,the mean method does not consider the distribution of the cluster point Clouds when estimating the centroid.The thesis proposed the centroid determination method based on Gaussian distribution（Gauss-CD）and the centroid determination method based on non-parametric kernel method（KernelCD）.At the same time,the thesis derived the Cramero Lower Bound（CRLB）for the both algorithms.Gauss-CD uses the Gaussian distribution to fit the distribution of point Clouds,and determines the position of centroid through maximum likelihood estimation.Experimental analysis shows that the point Clouds distribution fitted by the algorithm is not accurate enough,and its performance is similar to that of the mean method.To solve this problem,Kernel-CD used the non-parametric kernel method to fit the unknown distribution of the cluster point Clouds.The performance evaluation shows that,compared with Gauss-CD and mean method,the centroid estimated by Kernel-CD is closer to the true position of the target and has higher accuracy.（3）Combined with the extended target point Clouds obtained by the millimeter-wave radar in the road scene,the thesis designed the multi-target tracking algorithm based on the extended target.First,the algorithm uses ellipse threshold and group residual covariance matrix to design an adaptive correlation threshold;then,based on the centroid uncertainty and the radar measurement error,the adaptive covariance matrix of the measurement noise is estimated and applied to the Extended Kalman filter（EKF）and Unscented Kalman filter（UKF）;finally,combined with the logic method,the thesis designed the track management scheme.Experimental analysis shows that,the RMSE of the multi-target tracking algorithm based on extended target is less than 0.5 m,and the algorithm has good tracking performance in road scene.（4）Aiming at the problem of the robustness of tracking algorithms in road scenes,the thesis proposed the weighted robust EKF algorithm（ET-ROD-EKF）based on extended target.The algorithm weights the adaptive covariance matrix of the measurement noise,and the thesis verifies the tracking performance of the ET-ROD-EKF algorithm under Gaussian noise and Student’s t-distributed noise.The results show that ET-ROD-EKF is robust to outliers and effectively reduces the influence of outliers on the estimated position.