Infrared Human Target Detection And Action Recognion Based On Deep Learning

With the development of computers,new types of data,such as pictures and videos,continue to increase and become the main resources of mainstream hardware devices and Internet platforms.With huge potential,more and more machine vision researchers are focusing on how to make computers automatically and accurately understand human behavior in infrared video.In view of the above background,this paper investigates the current development status and difficulties of motion recognition,analyzes the development history and important algorithms of visible light motion recognition,and also investigates the progress of infrared motion recognition.Data set InfrAR,and the progress of the proposed recognition algorithms on this data set.Finally,the difficulty of motion recognition is analyzed by analyzing the spatial complexity,time difference,and computing resources.The text introduces the basic theories of deep learning,including the principles and structure of neural networks.It also analyzes several mainstream motion recognition algorithms based on deep learning.In the infrared human target detection task,this paper completes the entire process from model training to hardware implementation.In this paper,the infrared image data set of FLIR-ADAS is found and processed,and the dark-net training framework is used to train human targets in infrared images to obtain the network model parameters.Combined with Nvidia's open source optimization libraries TensorRT and DeepStream to accelerate the network model,real-time 25-frame infrared vehicle and pedestrian detection models are implemented on the Jetson Nano development board.This paper also analyzes the detection difference and real-time performance comparison between YOLOv3 and tiny-YOLOv3.In the motion recognition of infrared targets,this paper adopts the method of extracting motion frames based on NMPEG-4 compressed video,and directly extracts key frames,predicted frames and residual frames,And use deep learning networks ResNet152,ResNet18 to extract motion information from compressed content,use regularization technology and increase the number of video segments,can suppress the overfitting problem that is easy to occur in small data sets.Fusion weight experiments,repeated training and testing,find the optimal hyperparameters,and finally achieve 61.67%accuracy and the Mean Average Precision of 75.03%in the InfrAR infrared data set.It is proved that the motion recognition algorithm has certain practicability and applicability in the infrared field.Discuss and experiment with regularization weights,model fusion ratio,and the relationship between the number of video segments and the model,quantify the impact on the model,and better develop the performance of the algorithm.