Research On Human Fall Detection From Behavior Videos Based On Extreme Learning Machine

Fall detection is one of the main components in the smart home security monitoring.The research on fall detection has important significance for the elderly,especially the elderly living alone.With the development of machine vision technology and the wide application of video surveillance technology in life,the technology of fall detection based on video analysis provides a new security warning method for the elderly.Fall detection based on video analysis is realized through a process by use of the computer vision technologies,which include starget extracting,characteristics analyzing and the behavior classifying.After the analysis of these key technologies,a thorough research on the classification model is conducted based on the extreme learning machine and the algorithm of fall detection is constructed in this thesis.The details are given as follows:1.A classification model of extreme learning machine is proposed based on parameter optimization strategy.The classification accuracy of extreme learning machineis low due to the random assignment of the input weights and hidden layer offsets of the extreme learning machine.In order to improve the influence of the initial parameters,gravitational search algorithm is introduced.Meanwhile,the input weights and hidden layer offset parameters are regarded as the particles to be optimized by use the interaction between particles.At the same time,the particle velocity is limited to an interval in the process of the optimization.UCI database is used to verify the classification model.The experimental results show that the classification accuracy of the classification model based on the interval search optimization strategy is greatlyimproved.2.A fall detection algorithm based on clustering and parameter optimization ELM is proposed in this thesis to solve the problem about low classification accuracy of fall detection.Considering that there are some problems such as the incomplete description of the fall behavior during the process of the human target extracting and the failure to truly represent the behavioral process,the clustering strategy is introduced to cluster the texture features of the video frames,and the key frames of the human motion video are obtained,which are used to extract the target using the Gaussian mixture model.Moreover,an algorithm of fall detection is proposed based on all theabove analysis and extreme learning machine with parameter-optimized.SDUFall data set is used to test the algorithm.The experimental results show that the fall detection algorithm proposed in this thesis improves the accuracy of video fall detection.