Trajectory Association for Person Re-identification

Person re-identification (reID) aims at finding the same person in different camera views. In real-world scenarios, it is quite often that the suspect’s appearance is not known while the suspect’s escape route is known. This paper introduces a new person reID setting, where the query includes both the real suspect’s trajectory and several possible suspects. The goal is to identify the actual suspect and retrieve images of the real suspect. Prior work focuses on extracting pedestrians’ discriminative visual features or using spatial-temporal information while neglecting the importance of cross-camera trajectory information. Due to the spatial-temporal consistency, the trajectory and image complement each other and the trajectory is associated with the image data. Therefore, we consider retrieving the suspect’s image based on the trajectory and introducing a Hidden Markov Model based trajectory framework to jointly analyze image data and trajectory information. We evaluate our methods on two datasets containing person images and trajectory information, demonstrating our approach’s effectiveness.