Research On Fall Detection Algorithms In Indoor Monitoring Environment

With the acceleration of the aging of our country's population,home care has gradually become a trend.Elderly people are prone to falls when they are living at home.If they are not provided with timely assistance,they can cause physical and mental injury to the mild ones,and even endanger their lives in the severe ones.Therefore,research on methods that can detect the fall behavior of the elderly in time has great research value and social significance.With the popularity of surveillance equipment,especially the introduction of RGB-D cameras,computer vision-based fall behavior detection algorithms for the elderly have become a research focus.This article analyzes the research on indoor monitoring environment.The main work of the thesis is as follows:(1)An efficient fall detection method for RGB-D video based on temporal feature classification is proposed.Aiming at the characteristics of fall behavior and normal behavior,a fast fall behavior detection method for RGB-D video based on temporal feature classification is proposed.First of all,improvements have been made in the YOLOv3 target detection algorithm.For human target detection tasks,the backbone network structure has been simplified and the multi-scale prediction branch has been adjusted.An improved lightweight real-time human target detector has been proposed;secondly,improvements have been made.Feature extraction method of human fall behavior.Different from the common two-dimensional feature extraction ideas based on RGB images,it combines depth information to extend the feature dimension to three-dimensional space,and explores more discriminative and robust feature extraction methods;finally,use the depth that is good at sequence information processing The forest algorithm classifies the temporal motion characteristics to determine whether a fall has occurred.Experimental results show that the proposed fall behavior detection algorithm has achieved good performance on the UR Fall Dataset public data set and the RGB-D behavior data set independently collected by the laboratory.(2)A fall detection algorithm based on CNN-LSTM and attention mechanism is proposed.a fall behavior detection algorithm based on CNN-LSTM and attention mechanism is proposed via the depth image sequences of indoor human behavior.In the feature extraction network part,the convolutional attention module is introduced to improve the ability to express human visual appearance features;secondly,the time attention mechanism is introduced in the LSTM recurrent neural network part to help the network learn important parts of long-term information.Finally,the output of the network is classified by the softmax function to achieve end-to-end fall behavior detection.Experimental results show that the proposed fall behavior detection algorithm has achieved good detection results on several public data sets,with an accuracy rate of 97.07%.(3)A fall detection method based on neural network and multi-modal feature fusion.Most of the current fall detection methods are based on single modal data,and it is difficult to deal with the complexity and randomness of human behavior.Combining depth image and human skeleton information,we propose a fall detection algorithm based on deep neural network and multi-modal feature fusion.For the depth image sequences,the visual appearance features extracted by the convolutional neural network are then sent to the LSTM recurrent neural network to model the complex time dependence between the feature sequences;for the human skeleton sequences,the graph neural network is used to extract The topological features of the skeleton sequence,and the feature classification is realized through the softmax function.Finally,the classification results of the two modal features are post-fused to achieve fall detection.The experimental results show that the proposed method has improved robustness and accuracy compared with the single-modal fall behavior detection algorithm,and the accuracy rate reaches 99.21%.