Research On Multi-target Tracking Technology Based On Multimodal Fusion

Multi-target tracking technology is widely used in many fields such as security monitoring and automatic driving.It can determine the position and track of multiple targets in the detection range,and predict the motion information of the target.The traditional multi-target tracking technology is affected by the environment and the limitations of its own sensors,which is difficult to meet the tracking requirements in the actual scene.In this thesis,the multi-modal fusion method at decision level is used to track pedestrian targets by millimeter wave radar and monocular camera.The existing research is improved and optimized from three aspects: radar multi-target tracking,system track modeling and local track fusion,which improves the accuracy and robustness of multi-target tracking algorithm.The main research contents are as follows:1.A multi-target tracking algorithm based on millimeter wave radar is proposed to improve the accuracy of multi-target tracking.The algorithm introduces unbiased measurement conversion method in radar track modeling,so as to make full use of Doppler measurement information,and then cascade matching algorithm is used for data association of reliable track and temporary track respectively.Finally,the accurate tracking of multiple moving targets is realized.2.Based on the target track model in Deep SORT,a system track modeling method is proposed,which lays a foundation for track association and fusion of heterogeneous sensors.Aiming at the position information shared by radar and image,this method establishes the system track only containing the target position information,builds the radar track in the pixel coordinate system in the radar track,and separates the image track only containing the position information from the original image track,realizing the unification of the radar track,image track and system track state.3.A multi-target tracking algorithm based on multimodal fusion is proposed.The algorithm uses the spatial matching results of radar and image observation targets,takes the appearance matching degree as the measurement standard,makes track decision on the data association results of local track,and selects the correct data association results to update the local track.Finally,the sequential filtering fusion method is used to update the system track by using the state estimation results of local track.In the test based on the measured data set,the multi-target tracking accuracy is improved by 7.5 % compared with Deep SORT.By correcting the local track,this algorithm reduces the influence of the error results of the local track on the system track,and improves the accuracy of multitarget tracking by fusion tracking.