Keypoint Changes for Fast Human Activity Recognition

Abstract Human activity recognition has been an open problem in computer vision for almost 2 decades. During this time, there have many approaches proposed to solve problem, but very few managed it a way that is sufficiently computationally efficient real-time applications. Recently, changed, with keypoint-based methods demonstrating high degree of accuracy low computational cost. These take given image and return set joint locations each individual within image. In order achieve performance, sparse representation these features over time frame required classification. Previous achieved using reduced number keypoints, approach gives less robust the individual’s body pose may limit types can be detected. We present novel method reducing size feature set, by calculating Euclidian distance direction keypoint changes across frames. This allows meaningful individuals movements time. show achieves on par current state-of-the-art methods, while performance.