Search

This paper presents a novel method for movement disorder detection using a single Kinect sensor. Numerous recent studies have focused on gait analysis involving spatiotemporal parameters, such as the step length and gait cycle. However, these methods ignore significant kinematic information, e.g., leg swing characteristics. Therefore, in this paper, we 1) introduce a new concept called gait symmetry to measure the similarity of leg swing motions using the correlation among the angles formed by the hip, knee, and ankle joints; 2) extract the step length and gait cycle using a zero-crossing detection method; and 3) explore the K-means and Bayesian methods to conduct a combined gait analysis using the three extracted gait parameters. In addition, to reduce the effects of walking differences related to age and gender, we conduct a gait analysis of different age and gender groups. In our experiments, we collect a data set consisting of data from 120 participants (50 with disordered walking patterns and 70 with normal walking patterns). The experimental results show that 1) including the proposed gait symmetry parameter greatly improves detection accuracy, 2) joint utilization of the three gait parameters can enable significantly better performance, and 3) considering age and gender information can enable more robust detection.

This paper presents a sitting posture detection method based on a depth sensor. Effectively detecting bad sitting posture can help protect eyesight and prevent cervical spondylosis. Yet most current approaches conduct the sitting detection with traditional sensors, which leads to inferior performance. Therefore, this paper introduces a novel method based on a depth sensor, in which we 1) model the sitting posture by splitting body profile and locating six body joints, 2) extract two kinds of features: relative topological characters and local edge features, and 3) jointly fuse the two features for sitting posture detection. The proposed method is tested on our 3D sitting posture dataset, which includes 18 individuals, and totally 9 different sitting postures. The experimental results show that, using an adaptively fusion of the joints features, this method performs a high accuracy (more than 94%) and achieves a real-time detection (17 fps).