Human activity recognition based on LPA

Human activity recognition and fall detection have been popular research topics because of its wide area of application. Traditional activity recognition methods have complex feature extraction steps. We propose a new feature extraction method based on linear prediction analysis(LPA) to reduce computational complexity involved with engineering features. The feature extraction method we propose establishes a link between human activity and the signal system and regards acceleration signals as the output of the human activity. Using the relationship between the human activity and the output signal, linear predictive analysis can isolate information about human activity and transform it into a compact representation through linear prediction coefficients (LPC). In order to verify the effectiveness of the method, we design an activity recognition system based on linear prediction analysis and feature extraction. At the same time, we study the performance of the combination of linear prediction coefficients and time domain features. We use data from the public dataset SCUT-NAA, which contains ten different activities, and another public dataset, which records people falling. A random forest classification algorithm based on ensemble learning is used for activity recognition and fall detection. The results show that the combined vector of linear prediction coefficient and time domain activity amplitude feature obtained a 93% accuracy rate and the system evaluation index F1 of 0.92 on the SCUT-NAA dataset. Additionally, we achieved an accuracy rate of 97% in fall detection.