A Skeleton-based Action Recognition System for Medical Condition Detection

Progress in skeleton-based action recognition has enabled the contactless, ceaseless and portable surveillance on human daily behaviors which helps to reveal health hazards. Human action datasets like NTU RGB+D have provided medical-condition-related categories. Hence, in this paper, we propose an optimized view adaptive LSTM (VA-LSTM) with additional classification and regression subnetworks to such medical condition detection (MCD). We surpass previous works with a model accuracy of 79.6% (cross-subject) and 88.2% (cross-view). Besides, a dedicated hardware accelerator capable of multi-input processing is developed for this algorithm. Comparing to the algorithm implementations on general computation platforms such as Raspberry-Pi, Firefly-RK3399, and NVIDIA Titan X, the proposed accelerator achieves a less inference latency as well as a better energy-efficiency.