Research On Customer Tracking Algorithm For Supermarket Scene

With the development of artificial intelligence technology,artificial intelligence is more and more coming into people's lives,changing people's lives,such as face detection security at train stations,voice recognition input on mobile phones,etc.In the field of computer vision,Multi-Target Multi-Camera Tracking(MTMC)has become a research hotspot because of its potential commercial value.MTMC always points to determine the position of each person from a video stream taken from multiple cameras,and the resulting multi-camera trajectory can be included for visual surveillance,such as suspicious activity and anomaly detection,motion tracking,crowd behavior analysis,and cashier-less stores.This article mainly studied how to implement customer tracking in the context of supermarkets in conjunction with MTMC to help obtain user shopping data,such as activity trajectories in supermarkets,customer distribution heat maps,and product purchase order.Positioning to help businesses improve service quality and get maximum benefits.It mainly includes the following research points:1.Feature extraction by convolutional neural network.Research the existing pedestrian re-recognition algorithm with the best performance,propose a calibration learning algorithm that combines ranking-driven structural loss and classification loss,and a single-branch network structure,which was created and shot with 110 pedestrians under the overhead view.The pedestrian re-identification data set of the image validates the effectiveness of the proposed algorithm.2.2D image position data and 3D spatial position data based motion model modeling and multi-person tracking method research under single-camera vision.A motion model using two-dimensional 2D position data and shoulder slope and head 3D position data and shoulder-width data as inputs is proposed.In multi-person tracking under a single camera,the matching problem of the position data between the frames in the tracking is modeled as the best matching problem of resetting the bipartite graph according to the position matching cost.Multi-person tracking under a single camera.Aiming at the problem of invalid interference trajectory data caused by false detection and stationary targets under a single camera,an algorithm strategy of filtering using a convex hull algorithm was proposed.3.Research on trajectory matching methods between multiple cameras.Aiming at the problem of multi-camera and multi-person tracking,a strategy of hierarchically calculating trajectory data is proposed,that is,firstly calculating and acquiring the trajectory data of a single camera next,and then matching the trajectory data between multiple cameras to achieve each person's trajectory data in the entire scene.In the trajectory matching between multiple cameras,a trajectory matching method based on the feature similarity between trajectories,the position matching cost between trajectories,and the constraints of camera variable distribution is reduced,which reduces the amount of calculation and improves the system robustness..