Human pose tracking based on both generic and specific appearance models
Yao Lu, Ling Li, Patrick Peursum · 2012
Effective data association is essential for tracking human motion in monocular-video sequence. Data association using colour-based appearance models that are learned automatically and specific to the human being tracked has been shown to achieve good performance, but such specific appearance models can fail in cases where different parts have similar colour and often still require a prior training before the appearance can be learned. In this paper, a novel human tracking system is proposed that automatically extracts a specific appearance model and utilises this together with the initial generic appearance detector to estimate a human's pose in a video. No prior training or temporal smoothing is required. Experiments are conducted to compare the proposed approach against existing algorithms based only on specific appearances. Tracking is performed on several publicly available data sets to demonstrate that the approach works well without any training or tuning required, and results show that data association based on both generic and specific appearance models outperforms specific-only approaches.