Tracking Multiple Humans in Complex Situations
IEEE Transactions on Pattern Analysis and Machine Intelligence
| Hi-index | 0.01 |
Automatic detecting and tracking people from a stationary video camera is important for many applications. The problems are made difficult due to versatile human shape and appearance, persistent or temporary occlusion of multiple people and noise from various sources (e.g., shadow), which are commonplace in reality. We propose to tackle the challenges using applicable and general constraints in the form of models. In particular, we use a background appearance model and a camera model. We also use explicit human shape models as an entity for analysis both in segmentation and in global motion tracking, and use a 3D locomotion model to assist the estimation of articulated body postures. We present two approaches towards the goal of multi-human segmentation and global motion tracking. In the first approach, a simple shape model is used. The segmentation is done using direct image features. Multiple overlapping humans tracking is factored into tracking each one according to their depth order. The second approach follows a Bayesian framework. The optimal solutions for segmentation and tracking are defined explicitly as a Bayesian posterior probability in the joint-object space. The solution is computed by a Markov chain Monte Carlo-based method. The computation also takes advantages of domain knowledge as importance proposal probabilities to direct the Markov chain intelligently to obtain significantly faster convergence. The approach is more general and applies to the scenario where a large group of people has persistent occlusion. This approach has both the robustness and the optimality of the Bayesian formulation and the computational efficiency from the bottom-up processing. Although only applied to human segmentation and tracking, the proposed approach can be extended to general multi-object segmentation and tracking. We propose a tracking as recognition approach where the estimation of body postures is accomplished by recognizing the motion in a locomotion model. It results in robust performance in low-resolution data, with temporary occlusion and without interactive initialization.