基于百度飞桨的面向黑暗环境人员行为检测与身份识别

Human behavior detection and identification in dark environment based on Baidu Paddle

针对传统可见光在黑暗环境中难以实现人员行为检测与身份识别的问题,本文结合红外热成像技术基于百度飞桨深度学习框架研究了一种面向黑暗环境的人员行为检测与身份识别算法。首先经过实地采集,自主构建红外热成像人员行为数据集总计10900张9种行为类别以及双光人脸数据集总计3000张30位人员。针对行为检测方面,基于轻量化网络PP-LCNet改进YOLOv5骨干网络进行人员行为检测,大幅度减少模型参数并提高检测精度与推理速度。针对人脸识别方面,引入Cycle GAN算法改进Insight Face实现将红外人脸转化为可见光人脸进行身份识别,提高在黑暗环境下人脸识别准确率。最后实现红外人员行为检测网络与人脸识别网络的级联工作,在黑暗环境下可以实时行为检测与身份识别,具有很好的应用效果。实验结果表明,基于PPLCNet轻量化改进的YOLOv5相对于原网络模型参数减少56.4%,平均精度m AP由89.1%提高至94.7%,推理速度由68提高至101 fps;基于Cycle GAN算法改进Insight Face相对于原网络黑暗环境下识别准确率由84%提高至99%。

Aiming at the problem that traditional visible light is difficult to realize personnel behavior detection and identity recognition in dark environment,this paper combined with infrared thermal imaging technology to study an algorithm for personnel behavior detection and identity recognition in dark environment based on Baidu Paddle deep learning framework.First,after field collection,the behavioral dataset of infrared thermal imaging personnel totaled 10900 pieces of 9 behavior categories and the double-light face dataset totaled 3000 pieces of 30 personnel.In terms of behavior detection,the lightweight network PP-LCNet is used to improve the YOLOv5 backbone network for personnel behavior detection,reducing model parameters greatly and improving detection accuracy and reasoning speed.In terms of face recognition,CycleGAN algorithm is introduced to improve InsightFace to transform infrared faces into visible faces for identity recognition and improve face recognition accuracy in dark environments.Finally,the cascade of infrared human behavior detection network and face recognition network is realized,and real-time behavior detection and identity recognition can be achieved in the dark environment,which has a good application effect.The experimental results show that compared with the original network model,the parameters of YOLOv5 based on PPLCNet are reduced by 56.4%,the average precision mAP is increased from 89.1% to 94.7%,and the reasoning speed is increased from 68 to 101 fps.Based on CycleGAN algorithm,the recognition accuracy of InsightFace is improved from 84% to 99% in the dark environment of the original network.