Research On Image Target Tracking Based On Multi-source Information Fusion

In the field of security,products with wider monitoring range,higher detection accuracy and better tracking performance are often required.In this paper,based on the background of the actual project,studied the related technologies of target detection and target tracking based on millimeter wave radar and camera,and carried out the research on the problem of poor target detection accuracy and inaccurate tracking caused by low signal-to-noise ratio of long-range radar and small image target.In long-distance scenarios,the missed detection rate is difficult to fall below 5%,and the MOTA(Multiple Object Tracking Accuracy)is difficult to reach more than 50%,which cannot meet the actual project requirements.Therefore,the key points and innovations of this thesis are as follows:1.Aiming at the problem that it is difficult to detect long-distance "small" targets in the process of long-distance fusion processing,proposed a multi-source information fusion detection network based on millimeter-wave radar and images.Firstly,proposed an image cropping method based on least squares,which converts radar detection information into image coordinate information through coordinate transformation and generates a region of interest(ROI),which is cropped to increase the pixel ratio of the target in the image;Secondly,for the cropped image,this paper performs channel cropping on the basis of YOLOv5 and introduces the Ghost network to obtain an improved detection network YOLOv5 slimg,which achieves a balance between accuracy,calculation amount,and parameter amount;finally,the cropped image is sent to into the network for inspection.The experimental results show that: image cropping can make the average precision value m AP(mean Average Precision)of the detection algorithm reach 76.8%,and the performance is improved by nearly42.7%;the rate of fusion detection misses is 4.16%,which meets the practical application requirements and verifies the effectiveness of the fusion detection network.2.In order to solve the problem that IDSwitch(Identity Switches)frequently occurs due to association errors when tracking long-distance targets,proposed a multi-source information fusion tracking algorithm based on millimeter-wave radar and Deep Sort(Simple online and realtime tracking with a deep association metric)framework.Firstly,a data association algorithm is designed to realize the data association of multi-source information to the same target;Secondly,for the problem of detecting in formation loss when the target is occluded,the prediction information is used to complete the data of the missing part;Thirdly,designed a data association algorithm based on metric matrix fusion method.calculates the correlation coefficient according t o the covariance information of different targets,obtains the optimal weight of the corresponding measurement unit from the correlation coefficient,fuses the information of different measurement matrices and generates a fusion measurement matrix,so as t o improve the sensitivity of the algorithm to target actual distance;Finally the fusion metric matrix is used as the gating matrix to process the cosine matrix correspondingly to remove the impossible pairing results in the matching pairs,and improve the accuracy of data association.The experimental data show that the MOTA index of the fusion tracking method has reached 62.4%,the performance has been improved by 36.8%,the multi-object tracking accuracy MOTP(Multiple Object Tracking Precision)is 36.16 %,the number of IDSwitch has decreased by 30.4%,requirements and verifies the effectiveness of proposed algorithm.3.Combined with the project requirements,in order to verify the proposed algorithm and realize the visualization effect,designed a multi-source information fusion algorithm software platform,and implemented the fusion tracking processing algorithm.Firstly,according to the functional requirements of the system software,carried out the module division and multi-threading design;Secondly,designed the radar tracking algorithm,the algorithm and the fusion algorithm are integrated into the data processing module to expand the function of the system;Finally,used the Qt image application framework as a development tool for system software implementation.After the actual test,the software platform has completed the task of detecting and tracking the longdistance target area.