LoRa Posture Recognition System Based on Multi-Source Information Fusion

Abstract

Human body posture recognition is a popular research topic in the field of pattern recognition. At present, with advances in big data technology and data mining algorithms, long-term posture monitoring data of specific individuals or groups can be fully utilized. This data can be applied in the context of national sports statistics, health rehabilitation, and posture monitoring for the elderly. These applications involve large group and high-frequency usage; as a result, the system must have low energy consumption, a convenient carrying system, and low cost. Based on these requirements, this paper proposes a LoRa posture recognition system based on multi-source information fusion. LoRa WAN technology has the advantages of low cost and long transmission distance. A multi-sensor and LoRa terminal node are combined to collect human posture and are convenient to carry. The system includes a posture sensor module, wireless transmission module, and posture recognition module. The posture sensor module includes three types of sensors and a microprocessor for collecting posture information. The LoRa terminal node, LoRa gateway, and network server form a wireless transmission module for remotely transmitting data from the sensors. The function of the gesture recognition module includes pre-processing the sensor signals and outputting the pose categories based on the classification model. Aiming to address the disadvantage of the low LoRa transmission rate, this paper proposes a minimum redundancy maximum relevance sequential forward selection random forest (MRMR-SFS-RF) feature selection algorithm. This algorithm selects a small number of features that are most representative and outputs them for gesture recognition to increase the frequency of the feature output and reduce the energy consumption of the thermal node. To verify the performance of the proposed algorithm, feature data is extracted based on the proposed posture recognition system, and the accuracy of posture recognition is calculated. Finally, four key features are selected from 162 dimensional features. The experimental results demonstrate that the recognition accuracy for the four features is 98.9%, which achieves the goal of maintaining high classification accuracy with a small number of features.