Research And Implementation Of Fall Detection System Based On Feature Fusion And Improved Whale Optimization Algorithm

With the aging of the population,falls-related injuries and deaths among the elderly have become a major social concern.The "4-2-1" family model,which is unique to China,is increasing the number of empty nest families.Serious repercussions will follow if an elderly person falls and does not receive timely assistance.Fall detection and recognition have gotten considerable attention in recent years.However,the existing technique is inadequate at detecting similar activities,and misclassification is more frequent.This paper investigates the following aspects to reduce misclassification and increase the distinction between confusing activities.(1)To solve the problem of single feature for fall detection,this paper uses inertial guidance module to acquire 9-dimensional acceleration,angular velocity and angle.Moreover,features are extracted with three analysis methods:time domain,frequency domain and time-frequency to improve the feature diversity of the samples.The time domain method extracts features from the collected data such as maximum,mean and variance,which is less computationally intensive and easier to understand.The frequency domain method converts the signal from time to frequency using the Fast Fourier Transform(FFT)and extracts the maximum peak of the spectrum and other features.The Wavelet Packet Decomposition Method(WPDM)obtains spectral energy features from low and high frequency signals through local analysis.Combining the above three aspects of features,a total of 243-dimensional training sample is extracted.(2)Not all features in the artificial pose dataset are strongly correlated with the classification target.To improve the classification accuracy of the fall detection model,a feature selection(FS)method based on the Whale Optimization Algorithm(WOA)is used in this paper.Meanwhile,a multispiral Whale Optimization Algorithm(MSWOA)is proposed to address the problems of insufficient global search capability and easy to fall into local optimum.On the one hand,the algorithm uses the individual position and iteration characteristics to generate subpopulations adaptively,which can expand the search space and enhance the global exploration ability.On the other hand,individuals follow a bi-directional spiral path to continuously jump out of the local optimum and mine more promising regions near the subpopulation optimum to improve the performance of FS.Experimental results on the UCI benchmark datasets show that the feature subset obtained by MSWOA outperforms other comparative algorithms.Moreover,the F1-score result of the MSWOA-based fall detection model can achieve 95.66%.(3)The quality of the features can directly affect the classification results of the fall detection model.However,manual feature extraction relies on empirical knowledge of the field.The extracted features are mixed with human subjective factors,making it challenging to distinguish similar activities.Deep neural networks compensate for the lack of manual features by automatically mining the deep information of the samples to obtain the intrinsic features of the activities.Therefore,this paper proposes a fusion architecture obtains a fused feature dataset by combining deep features extracted from 1DCNN-LSTM networks and artificial features.The experimental results show that the fused features combine the advantages of both types of features and can obtain more effective feature information.Moreover,it has better recognition ability and the classification accuracy reaches 99.20%.