Combined Particle Filter and Its Application on Human Pose Estimation

Time series analysis has a wide range of applications in various domains. When addressing this task, tracking and identification are usually solved as two separate problems. However, it introduces a lot of computational redundancy and makes it difficult to guarantee synchronization and real-time performance. We propose a joint problem combining recognition and tracking and use particle filtering to solve the state estimation problem. However, the general particle filter cannot accurately estimate the state of the system in the time series state estimation task, because the system is time-varying and the prediction model of the particle filter is fixed. To address this issue, we assume that the system transition space is a set of finite prediction modes, and then propose a new Combined Particle Filter (CPF) framework that jointly achieves prediction mode recognition and state tracking. In the CPF, the prediction mode is regarded as a variable to be estimated, along with the system state variables. As a result, the resampled particle state set forms the global estimation of the system state, while the resampled mode variables indicate the optimal transition mode of the current frame. We construct an evaluation index system and conduct several evaluation tests to demonstrate the excellent performance of the CPF. Finally, we apply the CPF to the articulated human pose estimation task and obtain satisfactory results.