Research On Infrared Human Target Detection And Motion Understanding Technolog

Research on infrared personnel object detection technology based on deep learning is a necessary trend.This paper proposes a neural network algorithm for human body detection and an algorithm for human body key-point estimation and action understanding.The main contributions of this paper are as follows:(1)Based on the understanding and analysis of the actual requirements,an artificial intelligence algorithm design scheme of "object detection-key point estimation-action understanding" is proposed.(2)The overall design of infrared personnel detection algorithm based on deep learning is proposed.According to the actual needs of the task,a series of corresponding training datasets are proposed.These datasets construction strategy overcomes the time-consuming,and resource-scarce production difficulties of infrared datasets.(3)Based on the currently widely used object detection algorithm Faster-RCNN a novel object detection algorithm is proposed.Aiming at the problem of small infrared human objects,a cross-channel and hierarchical attention fusion feature extraction network is designed.In addition,a prime sample sampling attention guidance strategy is designed in the proposed network to improve the learning efficiency of the neural network.(4)A neural network algorithm for human body keypoint estimation and an action understanding neural network model based on the human body keypoint skeleton are designed.This paper uses the general visible human keypoint estimation algorithm Openpose as a template and makes specific improvements.The HRNet is used to improve the feature extraction architecture of Openpose.In terms of human action understanding,ST-GCN is used to realize human actions understanding.A training strategy is proposed to improve the learning efficiency of neural networks.(5)Systematically designed neural network training experiments,from the the two evaluation directions of horizontal and vertical,which quantitatively proves the superiority of the algorithm in this paper and the rationality of related improvements.