Part-aware trajectories association across non-overlapping uncalibrated cameras.

This paper focuses on the problem of multi-person tracking across non-overlapping uncalibrated cameras using data association method. The problem is extremely difficult as we have very limited cues to associate persons between cameras. To tackle the problem, our system consists of firstly building multiple trajectories from each camera independently, and then finding associations of trajectories between every two cameras of interest, where the later is the most challenging process. Our contributions are mainly two folds: First, we introduce a method to explore the human part configurations on every trajectory to describe the inter-camera spatial-temporal constraints for trajectories association. Second, we formulate trajectories association across non-overlapping cameras as a multi-class classification problem via the Markov Random Field (MRF) to effectively utilize domain priors such as group activity between persons. With the proposed part-aware correspondences and pair-wise group activity constraints of trajectories, we can achieve robust multi-person tracking. Experimental results on a benchmark dataset validates the effectiveness of our proposed approach. HighlightsWe introduce a method to explore the human part configurations aiming to describe the inter-camera spatial-temporal constraints.We formulate trajectories association as a multi-class classification problem via the Markov Random Field (MRF).We also proposed pair-wise group activity constraints of trajectories to achieve robust multi-person tracking.