Research And Implementation Of Fall Detection Algorithm Based On Smart Elderly Care

In recent years,China’s aging population and empty nest problem are becoming more and more serious,and elderly people living alone are often in unattended situations.According to the data falls have become the first cause of injury-related death for the elderly over 65 years old in China,and the implementation of real-time fall detection for the elderly has become an important safety guarantee and social need.With the development of artificial intelligence algorithms such as computer vision technology,sensor technology and deep learning,fall detection has become a research hotspot for intelligent elderly care.Current user fall detection algorithms are mainly based on computer vision and wearable sensor-based implementation.Among them,the fall detection algorithm based on computer vision uses a camera to obtain video information as data input,which has many problems,such as the limitation of complex backgrounds or occlusions that affect the recognition accuracy,poor robustness of unimodal detection and high complexity.To address the problem of poor single-modality robustness,this thesis uses wearable sensors to obtain data from another modality,extracts the corresponding feature information through a fall detection model,and finally performs multimodal fusion with video information to improve robustness.However,the fall detection algorithm based on wearable sensors still has many problems,such as feature extraction is not comprehensive and does not take into account the importance of different channel feature information,the complexity of the fall detection algorithm is high and the accuracy rate is not high.To address the shortcomings of existing research,this thesis conducts research in the following areas:(1)For the problems of low accuracy and high complexity of fall detection algorithms in computer vision,a fall detection model based on Feature Pyramids and Convolutional Block Attention Mechanism Networks(FPCN)is proposed.The model consists of a target detection module and a human pose estimation module,and the target detection module can pre-process the data and fall pre-determination,which reduces the complexity of the model.Meanwhile,a Convolutional Block Attention Module(CBAM)is added to the module to improve the recognition accuracy by focusing on the valid information in two different dimensions,channel and space,simultaneously.The human detection frame that has passed the target detection and fall pre-determination is cropped and fed into the human pose estimation module for subsequent key point labeling and fall detection,which reduces the time complexity of the model and improves the real-time performance.(2)To address the problem of poor unimodal robustness,this paper selects sensor data for fall detection in the second mode,which is sourced from wearable sensors.The fall detection algorithm based on wearable sensors is not comprehensive and complex in extracting features,so in this paper,we improve on DeepConvLSTM and propose a fall detection network(Squeeze-and-Excitation and Tensor-Train Decomposition Networks,STDN)model based on squeeze excitation network and tensor decomposition.The model uses Convolutional Neural Networks(CNN)and Long Short Term Memory Networks(LSTM)to extract features in the temporal and spatial dimensions,respectively,and fuses the two dimensions.The Squeeze and Excitation(SE)module is added to the model to rescale the features,weaken irrelevant features and strengthen important features to improve the recognition accuracy.Meanwhile,tensor decomposition is used to optimize the number of parameters and computation of the whole model to improve the real-time performance.(3)For the problem of fusion of two modalities,a decision layer fusion strategy is used for fall detection,in which the above-mentioned computer vision-based FPCN model and the wearable sensor-based STDN model are fused at the classification layer,and the results are output by a weighted summation method,and the final accuracy is improved.(4)All the above models were validated on the publicly available dataset UR Fall Detection Dataset(URFD)and custom data,and the effectiveness of the models was demonstrated through comprehensive experiments.The experimental results show that the computer vision-based fall detection model FPCN can effectively extract the person detection frame and make pre-determination by target detection,reduce the computational effort of the system,and can complete the fall detection efficiently.The fall detection model STDN based on wearable sensors can effectively improve the accuracy of fall detection by incorporating SE network,while using tensor decomposition to reduce the total number of parameters and computation of the whole model,and effectively reduce the complexity of the model.Multimodal fusion for fall detection further improves the accuracy of recognition and proves the effectiveness of multimodal fusion.