Behavior recognition system and method based on multi-position sensor feature fusion

Abstract

The invention belongs to the technical field of behavior recognition, and provides a behavior recognition system and method based on multi-position sensor feature fusion. The behavior recognition system based on multi-position sensor feature fusion comprises a data module and a model module, the data module stores and preprocesses collected original data to obtain action pictures for subsequent model training; the model module comprises a feature extraction sub-module and a classification sub-module, and a final prediction result is obtained. According to the method, spatial dependence among all data is abandoned, spatial dependence among same axial data of different sensors with stronger correlation is selected, three groups of action pictures are formed by extracting the data according to (x, y and z) three-axis directions and fusing the data, and feature fusion is performed through a secondary post-fusion model, so that the accuracy of the data fusion is improved. The time and space characteristics of the multi-dimensional time sequence data are effectively extracted, so that the aim of accurately identifying human behaviors is fulfilled.

Description

technical field [0001] The invention relates to the technical field of behavior recognition, in particular to a behavior recognition system and method based on multi-position sensor feature fusion. Background technique [0002] Human behavior is an indispensable part of people's daily life, and each behavior has its own special meaning and function. Accurately identify human behavior. Recognition is the basic function of many health monitoring, smart factories and sports scenes. It is important for promoting the field of artificial intelligence. progress is of great significance. At present, human behavior recognition is divided into behavior recognition based on video time series data and behavior recognition based on sensor time series data according to the data type. Compared with video data, wearable sensor data is not limited by the scene, and the storage and transmission are fast, making it easier to realize real-time recognition. For example, the elderly can wear sma...

AI Technical Summary

Problems solved by technology

The one-dimensional convolution kernel performs convolution on the time axis. When dealing with single-sensor and less-dimensional time-series data, it can effectively mine the time-dependency of time-series data. However, when facing multi-sensor multi-dimensional time-series data, one-dimensional convolution The kernel ignores the spatial dependence between different types of sensors and different axial sensor data, and cannot effectively extract the spatial features of multi-dimensional time series data

The two-dimensional convolution kernel can solve the above problems, but when constructing two-dimensional spatio-temporal data input, it simply splices the axial data without considering the spatio-temporal dependence between the axial data

However, the two-level post-fusion strategy still simply sends the sensor information into the convolutional neural network in terms of feature extraction, and fails to design an effective feature extraction method to further extract the spatial features of different sensor information.

Images