Learning deep features for online person tracking using non-overlapping cameras: A survey

摘要

Target-agnostic person tracking and re-identification across multiple non-overlapping cameras is an open vision problem. It is the task of maintaining the correct identity of people at different time instances and possibly different cameras. This study focuses on existing algorithms that facilitate online person tracking by using discriminative spatio-temporal features from video data, and presents the open issues and future research directions. The initial take on the problem introduces person tracking as a pure association problem, where the influence of human appearance, biometric and location information on re-identification are addressed explicitly. These constraints are modeled and used to understand and associate detections in real world environments. Next, a spatio-temporal model using LSTM networks for propagating associations and recovering from errors by taking advantage of the spatial and temporal information in videos is described. The spatio-temporal context indicates a way for discriminative appearance learning. The novelty of the mentioned approaches is that they do not require to learn target-specific appearance models and collect samples to distinguish different people from each other. The methods are evaluated on large-scale tracking datasets. State-of-the-art performance is achieved using motion metadata such as person bounding box and camera number, and shows better associations for the challenging exit-entry cases.