Body-part tracking from partial-view depth data

Abstract

This paper presents a high-accuracy body-part tracking algorithm, capable of achieving efficient human motion analysis from partial view depth-data, suitable for deployment in real-life applications. The algorithm uses a consumer-grade depth camera for data input and combines a discriminative body part estimator along with a generative tracker, utilizing a realistic human body model, in order to track individual body limbs in short camera-distance, partial-view scenarios. Additionally, a shape adaptation feature is also introduced in order to further morph the human model based on the observations. The implementation is tested in a lower-body limbs tracking scenario, achieving promising accuracy and performance on consumer-grade hardware. Moreover, a lower-body motion dataset is also provided, consisting of 16 real-world sequences using automatic ground-truth annotations from a commercial motion capture system.