A Real-Time Fall Classification Model Based on Frame Series Motion Deformation

Abstract

Fall is common among the elderly and patients with severe health conditions. It can be life threatening, especially if the person does not received the required help to recover from fall on time. Therefore, an automatic real-time fall detection system is very desirable and has been the focus of research in the last couple of years. Traditional computer vision (CV) based fall detection systems require less infrastructure, is cheaper and can be more efficient comparing to systems based on wearable sensors. The robustness and efficiency of CV techniques depend on the extracted feature set from the surveillance video sequences. Generally, the efficiency and accuracy of more recent learning based CV systems rely heavily on the statistical characteristics of training dataset. Acquiring a balanced, comprehensive and representative training data that covers all the necessary aspects of the problem, including viewing direction of the camera and illumination condition of the environment is quite challenging. The problem would be even more serious when the training dataset does not have representative features as the surveillance area. In this paper, we propose a robust, real-time, CV based fall detection technique that can work in different settings. The propose system requires only a single affordable RGB camera. The proposed method works at frame level and only uses two significant feature points for classification, therefore occlusion would not influence the system. We performed experiments on different publicly available datasets such as le2i, UR and multiple camera fall detection datasets. The result shows that the proposed technique can distinguish fall from everyday activities, e.g., sitting down and sleeping and has a higher accuracy, recall and specificity comparing to other methods. The proposed method performs well in indoor environments with different lighting conditions and different viewing directions of the camera.