Real-time Face Detection And Tracking Based On Video Stream

In the field of computer vision,face detection and tracking is a basic technology which is the first step before analyzing and understanding human faces in video images.Due to the variability of the face and some uncontrollable interference factors of the external environment(e.g.,face posture change,complex background,face occlusion,scale change),it is still a challenging topic to design real-time face detection and tracking algorithms with good performance.This thesis analyzes the latest face detection and face tracking algorithms and makes a certain degree of improvement,which improves the performance of the algorithms.In the part of face detection,the STFD(Scale-invariant Tiny Face Detector)detection algorithm is proposed.Firstly,the feature extraction module adopts a full convolutional neural network,which consists of the first five-layer network structure of VGG16 and an additional three-layer network structure.This module structure will automatically learn the complex features of the image,including low-level spatial detail information and high-level rich semantic information.Then using six scale feature maps to detect faces of different scales to solve the multi-scale detection problem of the face,in which large scale feature maps(lower convolutional feature maps)will detect small scale faces,and small scale feature maps(high level convolutional feature maps)will detect large scale faces.Finally,for the problem of missed detection of small faces and outlier faces,the method of increasing the anchor density on the feature map of low-level convolution is designed to increase the number of anchor of small faces and outlier faces,which improves the detection rate.The experimental results on the PASCAL face dataset,FDDB and WIDER FACE validation dataset show that the STFD face detection algorithm alleviates the missed detection of small faces and outlier faces to some extent.In the part of face tracking,a face tracking algorithm that combines attention and feature fusion is proposed.This thesis analyzes the shortcomings of the full convolutional siamese network for tracking tasks and improves it.Firstly,combining the first frame and the previous frame of the current frame as a target template to update the target template in real time.Secondly,the features of the multiple convolution layers are extracted,which aims to merge the apparent information and semantic information of the object.Finally,the channel attention mechanism is combined with the multi-layer convolution features of the target template,which gives higher weight to the channel features that have a greater impact on the tracking object,and improves the discriminative power of the target template features.Compared with several classic tracking algorithms,the proposed tracking algorithm achieves a more competitive effect on the basis of real-time performance.