SKIP: Accurate Fall Detection Based on Skeleton Keypoint Association and Critical Feature Perception

As deep learning technology advances, human fall detection (HFD) leveraging convolutional neural networks (CNNs) has recently garnered significant interest within the research community. However, most existing works ignore the cross-frame association of skeleton keypoints and aggregation of feature representations. To address this, we first introduce an image preprocessing (IPP) module, which enhances the foreground and weakens the background. Diverging from common practices that employ the off-the-shelf detector for target position estimation, our skeleton keypoint detection and association (SKDA) module is designed to detect and cross-frame associate the skeleton keypoints with high affinity. This design reduces the misleading impact of ambiguous detections and ensures the continuity of long-range trajectories. Further, our critical feature perception (CFP) module is crafted to help the model learn more discriminative feature representations for human activity classification. Incorporating these components mentioned above, we introduce SKIP, a novel human fall detection approach, showcasing improved detection precision. Evaluations on the publicly available telecommunication system team v2 (TSTv2) and self-build datasets show SKIP’s superior performance.

I. Introduction

Human fall detection (HFD) aims to continuously and precisely identify fall events executed by humans, which is both fundamental and challenging. Research findings indicate that HFD has a profound and extensive influence in various domains, such as healthcare [1], security surveillance [2], [3], and pose recovery [4], [5]. Although the fall detection community witnessed encouraging progress in the past years, robust fall detection in complicated scenes is still struggling to cope with diverse types of interference, such as posture variation, dim light, and background clutter. For example, fall activities undertaken by humans exhibit a wide range of variations, encompassing the fall’s direction and velocity, along with the body’s posture throughout the descent. Nevertheless, the unpredictability of human movements and the changing environment pose challenges to these detection systems’ accuracy in predicting falls. Furthermore, the effectiveness of many detection methods largely depends on the quality of the learned feature representations, rendering them susceptible to cluttered backgrounds. Although several attempts have been made to tackle this problem, they typically struggle when similar individuals interact frequently. To date, reliable and performing solutions for daily human falls are being extensively studied.