Research Of Human Activity Recognition Algorithm Based On UWB And Support Vector Machine

Due to the strong multipath resolution, penetration ability and low power consumption characteristics, Ultra-wideband (UWB) wireless communication technology is widely used in obstacle detection and object recognition and other fields. Meanwhile, with the development of human-computer interaction, there are more and more researches focus on human activity recognition (HAR). A novel HAR method has been put forward by combining machine learning with UWB signal sensing information. This paper contains the following three aspects:UWB sensor information feature extraction, improved Chaos Adaptive Genetic Algorithm (ICAGA) and verification platform of HAR, the main work is as follows:In order to extract the feature of the activity effectively, a gesture feature extraction method based on wavelet packet decomposition is proposed. First, each frequency component is achieved by using wavelet packet decomposition, then the energy of each sub-band can be calculated and the normalized wavelet packet energy distribution can be acquired finally. The results show that the wavelet packet energy distribution feature has good discriminability.In activity recognition, the performance of SVM has a greater impact on the selection of parameters. In this paper, the improved chaos adaptive genetic algorithm (ICAGA) is selected to optimize the parameters of SVM. Moreover, ICAGA adopts the dynamic cross rate and mutation rate according to the group fitness, thus effectively avoiding the disadvantages of standard GA, such as premature convergence and low robustness. Finally, ICAGA is applied to optimize the HAR model, and the actual dataset is collected by UWB apparatus. The results show that the proposed algorithm has good globe search capability and fast convergence, can improve the recognition accuracy effectively.Based on the analysis aforementioned, an algorithm verification platform for HAR is build based on GUI design in MATLAB. The UWB transceiver parameters setting and signal pre-processing, feature extraction and model training are integrated to the platform.Finally, the research works of the whole paper is summarized, and some valued suggestions about the activity recognition are discussed.