Deep Network-based Slow Feature Analysis for Human Fall Detection

Abstract

One of the most concerning safety hazards for elderly people is abnormal falls in public places. Vision-based fall detection using ambient cameras is a popular non-intrusive solution. Recent research uses Slow Feature Analysis (SFA), which can learn the slow invariant varying shape features obtained from input signals and is efficient. Another recent famous approach in motion detection is deep learning. However, the fall event in actual cases is diverse, resulting in complications in the detection task. Additionally, it is difficult to acquire fall-related data; hence, simulation is done on fall events to generate a training dataset, resulting in smaller data. Considering these complications, we have presented a novel method by combining SFA, deep learning models, namely Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM), and rule-base. CNN is used to extract the object region, thereby reducing the region of interest (RoI). Two shape features, such as aspect ratio and area of RoI are considered as input to the LSTM for retrieving the temporal information which is further used for rule generation, thereby increasing the detection accuracy. The efficacy of the proposed method for various features, such as aspect ratio, area, and aspect r a tio+area is demonstrated over the UR Fall data with an accuracy of 95.2%, 93.8%, and 96.36%, respectively.