Research On Online Counting Of Moving Objects Based On 3D Multi-target Trackin

In the production and packaging of industrial assembly line products,it is often necessary to inspect their quantities in real time,especially for small-sized targets such as capacitors and tablets,which are in high demand.Although currently there are methods based on two-dimensional image processing that can realize counting,the methods of two-dimensional vision are very limited in practical use,and it is difficult to cope with problems such as mutual occlusion of objects in unstructured scenes and complex lighting environment.Therefore,this paper studies and designs an online counting method for moving objects based on three-dimensional multiple object tracking,which provides technical support for real-time automatic counting of small objects such as capacitors and tablets.The main research work is as follows:(1)Preprocessing of raw point cloud data.Firstly,combined with the characteristics of depth camera point cloud data generation,a three-dimensional background difference method that can be used to detect three-dimensional moving objects is implemented.Then,Pass-through filtering and voxel filtering are used to remove the noise and redundancy of point cloud data respectively,Finally,the discrete point topology is established by using Kd-tree,which provides data preparation for target segmentation and tracking tasks.(2)Aiming at the difficulty of point cloud segmentation caused by mutual occlusion and collision between tantalum capacitors and tablets in scattered scenes,a segmentation method based on LCCP and adaptive regional growth threshold is proposed.Firstly,the point cloud clusters are over-segmented into super voxels,and the super voxels are segmented based on convexity.Then the edge contours of the under-segmented point cloud clusters are rejected based on the average normal angle of the clusters and the point density feature,and finally the improved regional growth algorithm with adaptive growth threshold is used to complete the final segmentation.The effectiveness and accuracy of the proposed target segmentation algorithm are verified by experiments.(3)A self-correcting data association method based on LAPJV algorithm was proposed to solve the problem of continuous tracking errors caused the temporary disappearance of targets and the occupation of trajectories.This method combines Euclidean distance and target motion vectors to establish a matching cost matrix to enhance the correlation degree between the trajectory prediction centroid and the detection target.The abnormal matching matrix was established for secondary data association to correct the wrong matching results.It is experimentally demonstrated that the tracking accuracy of the multi-object tracking method based on this data association designed in this paper reaches 96%,the tracking accuracy is less than 0.7mm,and it has certain robustness.(4)The object counting system is designed and implemented to provide a user-friendly human-computer interface for performing counting operations.The tracking and counting results are displayed in the visualization window in real time,which is convenient for users to read and check the counting results.In this paper,in order to realize the counting of capacitors and tablets in scattered motion scenes,an online counting method of moving objects based on three-dimensional multi-object tracking is proposed and implemented by taking the proposed or improved target point cloud segmentation and three-dimensional multi-object tracking algorithm.It is verified that this system designed in this paper can accurately complete the real-time online counting of tantalum capacitors and tablets,and the processing time of single frame data is less than 200 ms,which basically meets the real-time requirement.