Research On Multi-camera Multi-Target Tracking Towards Warehouse Scene

In the modern logistics warehousing scenario,With the increasing express parcel throughput and the enhancement of personnel mobility,strengthening the safety of goods and improving logistics efficiency are two important goals of warehouse management system.Leveraging the full-coverage cameras in the warehouse scene and buildding an intelligent monitoring system is an important cornerstone for realizing the above goals.In this paper,we design and implement a personnel global locating tracking system for warehouse scene.By using the cameras evenly covered in the warehouse scene,we can realise Multi-Personnel Multi-Camera Tracking and visualize the tracking process.The main work content and innovation points of this paper are as follows:(1)In order to solve the problem of occlusion detection in warehouse scene,we design an occlusion pedestrian detection algorithm.Specifically,we integrat an occlusion probability prediction network into RetinaNet,and proposed Separate NMS(S-NMS)screening mechanism to improve the detection accuracy of occlusion personnel.We build a detection dataset towards the warehouse scene,the experimental results tested on our dataset show that our algorithm outperforms RetinaNet,achieves 98.2%AP and 1.1%higher than RetenaNet.(2)In order to solve the problem of tracking multi-pedestrian with uniformed clothing and realise cross-camera tracking in the warehouse scene,we combined the multi-target tracking algorithm with pedestrian re-recognition technology and propose a siamese network based multi-target tracking algorithm.We build a cross-camera pedestrian dataset towards the warehouse scene,and the experimental results tested on our dataset show that our algorithm can greatly enhance the performance of feature extraction and marvelously increase the Rank-1 accuracy of cross-camera tracking,achieves 80.8%Rank-1 and 13.8%higher than original network.(3)We build an UI operating platform for Multi-Pedestrian Multi-Camera Tracking towards warehouse scene,which consists of a tracking-pool updating mechanism and a multi-pedestrian tracking framework.The tracking-pool updating mechanism can unify the camera network and realise cross-camera tracking.In the multi-pedestrian tracking framework,we consider if personnel is occluded and optimize the matching method to enhance the robustness of tracking.Finally,we make use of the camera calibration technique to visualise the trajectories on a unified warehouse map.