| Smart-Retail,as a popular direction in the commercial application of artificial intelligence,is considered to be the future of the retail industry.The Smart-Retail is the in-depth mining and analysis of key data such as customer movement trajectories and distribution heat maps to predict customer consumption trends and guide the production and manufacturing of the supply chain.In the field of smart retail,Multi-Target Multi-Camera tracking technology(MTMC)based on computer vision is used to obtain the movement trajectory data of customer targets in the scene.Currently,there are many difficulties and challenges in the MTMC task based on Smart-Retail scenarios.The first is that the environment of the use scene is complex,and the phenomenon of occlusion and overlap between targets is frequent,which increases the difficulty of target tracking.Secondly,the computing power of equipment is limited,and the actual terminal deployment is mostly low-cost and small computing power equipment,which requires the tracking algorithm to be efficient.In response to the above problems,this article first studies and designs a lightweight target detection model,and uses the detection and tracking fusion framework to further improve the tracking accuracy and operating efficiency of the algorithm.Finally,an efficient pedestrian re-recognition model is designed based on the feature aggregation network to complete the target cross-camera tracking.The lightweight MTMC tracking method proposed in this article can not only achieve good tracking results in SmartRetail scenarios,but also meet the needs of computing power constraints.The main work and innovations of this paper are as follows:1.Based on the MOT tracking mode of detection,an improved lightweight target detection model is proposed.Target detection is a step that takes up a lot of computing resources in a MOT tracking algorithm.In order to improve the efficiency of target detection,this paper optimizes its design from both backbone and detection head.In the feature extraction part,the model size is compressed by modifying the size of the sensing field and balancing the depth and shallow feature channels.After that,a CEM and a SAM are added to the back-end instance detection part to optimize feature output and improve detection accuracy.By compressing and optimizing the front and back ends of the detection network at the same time,the target detection model is superior to other mainstream methods in terms of detection effect and inference speed.2.Based on the fusion framework with detection and tracking,the detection module and the tracking module are embedded in backbone,so the efficiency of the MOT tracking system is improved.Using the structural characteristics of the two-stage target detection network,this paper uses a fusion framework to get rid of the tracking paradigm of frameby-frame detection and data association in conventional MOT tracking algorithms.The data association module and the detection module are integrated to improve tracking efficiency.At the same time,this paper also designs a motion model prediction branch network that can optimize the performance of MOT tracking performance in complex environments.3.Research and design an efficient person Re-Identification network.In order to meet the needs of low cost and low computing power equipment,we research and design an efficient person ReID model.By analyzing the shortcomings of the mainstream backbone network,this paper adopts a multi-layer feature aggregation backbone,and using Circle loss with better convergence,so that the ReID model can output feature vectors with higher discrimination.Compared with other models,our method has higher accuracy and faster inference speed. |
