Human behavior micro-Doppler classification and recognition method based on migration deep neural network

Abstract

The invention discloses a human behavior micro-Doppler recognition method based on a deep migration network, which utilizes ImageNet and other large-scale open source natural image databases to pre-train the deep neural network, optimize the network weights of the deep network, and then perform different behavioral Micro-Doppler spectrum is used for supervised learning. During the learning process, the network weights of the convolutional network are frozen, and only the fully connected layer of the deep network is trained. The training cost function is the combination of the Softmax function and the natural image and micro-Doppler spectrum. The sum of entropy differences, the trained neural network can be used to effectively distinguish the micro-Doppler time spectrum of different behaviors, so as to realize the classification and recognition of human behaviors. The invention adopts migration learning to optimize the network weight of the deep convolutional network, which can effectively reduce the requirement of the deep neural network for the training data set and improve the performance of classification and recognition.

Description

technical field [0001] The invention relates to the field of radar target micro-Doppler effect analysis, in particular to a human behavior micro-Doppler classification and recognition method based on migration deep neural network. Background technique [0002] Radar can locate and identify targets by emitting and receiving electromagnetic waves. For a moving target, the frequency of the transmitted and received echoes will change with the moving speed of the target, which is the Doppler effect. Since the movement of the human body is a non-rigid body movement, each part of the body has its own regular movement during the movement process. This frequency modulation of each limb added to the human body is called the micro-Doppler effect. The micro-Doppler effect is not sensitive to distance, light conditions and background complexity, and micro-Doppler features corresponding to different behaviors can be extracted and analyzed for analysis and estimation of human motion chara...

AI Technical Summary

Problems solved by technology

Due to the particularity of the composition of human targets, the instantaneous feature intensity of the whole human body (trunk and head) in radar echoes is much greater than that of limbs, so it is often difficult to directly extract the instantaneous features of limbs in radar echoes

Images