A Novel Indoor Human Fall Detection Method Based on an End-to-End Neural Network and Bagged Tree Classifier
Fayu Wang, Jianyang Liu, Guangdi Hu · 2019
In this paper, a novel indoor human fall detection method based on an end-to-end neural network and bagged tree classifier is proposed. This method is different from the conventional subjective feature extraction based fall detection algorithm, which combines the advantages of above-mentioned two algorithms and converts the fall detection problem into target tracking, features extraction and posture recognition problem. Firstly, a lightweight neural network is utilized to extract informative features from the video in the initial tracking process. Secondly, the extracted features are processed by discriminant correlation filter to get the correlation information of two adjacent frames which is used in the following process of inverse fast Fourier transformation to obtain the response map. The position of the predicted target is then updated by getting the maximum value of the response map. Furthermore, the position predicted in the previous frame is cropped and expanded as the input in the prediction of the next frame to achieve online tracking and feature updating. Finally, the updated features are used as the input for training bagged tree classifier. After that, posture recognition task is carried out by using the trained bagged tree classifier. The experimental results demonstrate that the proposed method performs well and achieves reasonably accuracy.