A multi-size input HAR algorithm based on an improved LSTM-CNN

Abstract

In order to solve the problem that a traditional machine learning algorithm has difficulty in feature extraction and confusion of similar actions to classification and recognition, the invention provides a multi-size input HAR algorithm based on an improved LSTM-CNN. According to the algorithm, an optimized CNN convolutional neural network and a multi-layer bidirectional dynamic LSTM long-short time memory network are fused to identify HAR human body behavior activities. The method comprises multi-size input and model classification and recognition, the multi-size inpu subjects the data to theoperations of noise reduction, mean value-variance normalization, zero filling and the like to generate two different dimensions of data as model input, and an optimized CNN network and a multi-layerbidirectional dynamic LSTM network are fused for classification and recognition through model classification and recognition. According to the method, the optimized convolutional neural network and the multi-layer bidirectional dynamic long-short-term memory network are fused to construct the classifier, so that the method has good expansibility and robustness, and can realize high-precision human behavior activity recognition.

Description

technical field [0001] The invention belongs to the field of pattern recognition and artificial intelligence, and in particular relates to a multi-size input HAR algorithm based on an improved LSTM-CNN. Background technique [0002] Sensor-based human behavior recognition is an emerging research direction in the field of human behavior recognition. Judging human behavior status by analyzing the acquired human behavior information has high research value and broad application fields. Human behavior recognition has two directions, sensor-based and visual-based. Vision-based human action recognition processes and analyzes the original images collected by the camera, learns and understands the actions and behaviors in them, and establishes the mapping relationship between video content and action type descriptions, while vision-based HAR (Human Activity Recognition ) is limited by specific scenarios, time, etc., which will directly affect the correctness and robustness of HAR a...

AI Technical Summary

Problems solved by technology

[0004]Traditional classification and recognition algorithms use manual extraction of effective feature values ​​for classification and recognition. For actions with great similarity, feature extraction becomes particularly difficult

Images