A low-cost framework for the recognition of human motion gait phases and patterns based on multi-source perception fusion

Abstract

Due to the increasing resolution and usage requirements of sensors and the increasing computational cost of algorithms, the application of existing human motion pattern and phase recognition systems in wearable systems is limited. This paper proposes a recognition system framework based on multi-source information fusion for the recognition of human gait patterns and phases. The framework utilizes multi-channel commercial low-cost sensors to obtain human motion information. By applying data fusion methods and preprocessing algorithms, the framework can uniformly process raw signals of uncertain dimensions from an uncertain number of sensors into a valid recognition vector with fixed dimensions. The obtained recognition vector can be used with commonly applied gait recognition algorithms to reduce the computational cost. Support vector machines, Backpropagation neural networks, AlexNet, and LeNet5 algorithms are used to evaluate the performance of the proposed gait recognition framework in recognizing gait phases and patterns. The experimental results show that the four algorithms using fusion signal can achieve a recognition accuracy of 97.7% for gait phases and an average recognition accuracy of over 99.2% for gait patterns, which proves the effectiveness of the proposed framework.