基于Retinaface与KCF人脸目标检测与跟踪算法研究

Research on Face Target Detection and Tracking Algorithm Based on RetinaFace and KCF

虽然随着深度学习的快速发展,在人脸的检测跟踪领域取得了一定的研究成果,在理论上不断进步创新,但在实际应用中仍然存在很多问题,比如人脸目标跟踪的速度慢的问题、光照强度不断变化下以及遮挡情况下导致检测与跟踪的最终效果与预期相差较大的问题等。为了解决此类问题,该文将深度神经网络(Retinaface)和核相关滤波器(KCF,KernelizedCorrelationFilters)进行结合,提出一种新的人脸跟踪算法Retinaface-KCF,通过Retinaface对视频流中的人脸进行检测,将检测到的人脸坐标信息传递给KCF进行跟踪。在跟踪过程中,设定两种异常情况处理方法,第一种是异常峰值,通过实时监测响应峰值,当出现异常峰值时,则判定为跟踪目标异常,通过Retinaface重新进行检测。第二种是时间阈值,通过设定目标跟踪重新检测阈值,当目标跟踪超过阈值时间,同样通过Retinaface重新进行检测。经过实验表明,该方法在保持较高的准确率情况下,可以达到平均150FPS的速率。

While with the rapid development of deep learning,in the field of human face detection tracking has obtained certain research results,progress innovation in theory,but still exist many problems in practical app lications,such as face target tracking problem of slow,light intensity and shade under the changing circum stances cause large difference detection and tracking of the real and expected effects of problems.In order to solve this problem,a new face tracking algorithm,RetinaFace-KCF,is proposed by combining deep neural network(RetinaFace)and Kernelized Correlation Filters(KCF).The face in video stream is detected by Retina Face,and the detected face coordinate information is transmitted to KCF for tracking.In the tracking process,two abnormal situation processing methods are set.The first one is abnormal peak,and the response peak is monitored in real time.When the abnormal peak occurs,the tracking target is determined to be abnormal,and Retinaface is used for re-detection.The second is the time threshold,which is re-detected by setting the target tracking.When the target tracking exceeds the threshold time,Retinaface is also used for re-detection.The experimental results show that the method can achieve an average rate of 150 FPS while maintaining a high accuracy.