Video Face Detection Based on Improved SSD Model and Target Tracking Algorithm

Abstract

Video face detection technology has a wide range of applications, such as video surveillance, image retrieval, and human-computer interaction. However, face detection always has some uncontrollable interference factors in the video sequence, such as changes in lighting, complex backgrounds, and face changes in scale and occlusion conditions, etc. Therefore, this paper introduces deep learning theory and combines the continuity characteristics of video sequences to make related research on video face detection algorithms based on deep learning. First, this algorithm uses the residual network as the basic network of the Single Shot MultiBox Detector (SSD) target detection network model and trains a Rest-SSD face detection model to detect faces. Experimental results show that the method can achieve real-time detection and improve the accuracy of video face detection, which is required for face detection in a video. Then we based on the continuity characteristics of video sequences. This paper proposes a video face detection method based on the training of the Rest-SSD face detection model. The method first uses kernel correlation filtering to track consecutive n frames according to the detection results, sets weights on the confidence of the n frames of tracking results, uses the weighted average method to calculate the best tracking result, and then sets the best tracking result confidence and the current frame sets the appropriate weights for the confidence of the detection result for fusion, thereby improving the video face detection accuracy.