Camera based Activity Recognition for Assisted Living Applications
Saroj Kumar, Ayesha Choudhary · 2022
Activity recognition in the assisted living environment could detect dangerous and anomalous activities. One such activity is the act of falling, which is very dangerous and can cause injuries and death. Falling is a leading cause of unintentional injury and death worldwide. The unintentional deaths caused by falls can be prevented to a great extent by its early detection and providing subsequent medical help. Although several fall detection systems exist, they lack speed, most of which are slow in inference time. We propose a single-shot fall detection model as part of camera-based activity recognition for assisted living applications which is fast and accurate. The proposed fall detection model uses a deep learning-based object detection framework called YOLOv5. Instead of following a two-step process for fall detection, namely, subject of interest detection (object detection) and activity recognition (fall activity recognition), we detect falls in a single stage by virtue of object detection and activity recognition merged into a one-step process. This approach introduces novelty by using an object detection framework for activity recognition and fall detection. The proposed model is implemented in the PyTorch framework. We train different variants of YOLOv5 (small, medium, large) differentiated by their depth. The best result obtained shows an inference speed of 32 frames per second. This fall detection model detects falls accurately in challenging conditions with noise and background clutter.