Research On The Face Tracking Algorithm Under Complex Illumination

With the rapid development of artificial intelligence, pattern recognition, digital image processing, as well as the gradual implementation of the Sky net project, intelligent video surveillance gets more and more attention, and has developed into a widely used comprehensive technology. Automatic detection and tracking of video human face region is one of the key technologies of intelligent monitoring system. With further research, scholars have proposed many tracking algorithms.In practical applications, if the ambient illumination changes during the tracking, especially when there is a sidelight, the face area tracking lead to inaccuracy or failure, and complicated illumination enhance algorithm can not guarantee the real-time tracking. Therefore, we studied the real-time human face area tracking algorithm in complex illumination conditions:In the aspect of face detection, we proposed a new automatic face detection algorithm for the first frame of the given video. We proposed a face image preprocessing algorithm which can reflect the original feature of human facial organs by eliminating the effect of illumination on different human facial organs. The algorithm effectively reduces the influence of illumination on the human eye positioning. Criteria for the human eye is then given based on a priori knowledge of the geometry location of the person’s face, and then we can determine the position of eyes by automatically adjusts the segmentation after estimating the image segmentation threshold range. Finally, we can find the location of human face in the images based on the location of the eyes.Since the processing time of the pre-treatment methods can not meet the requirements of real-time video tracking, we presents another way of image preprocessing which named the local adaptive image gamma correction algorithm to solve the unevenness of illumination of subsequent frames of video after the first frame and meet the requirements of real-time video tracking, the algorithm can adaptively correct the gray-scale of the image without knowing the image’s illumination. The basic principle of local adaptive image gamma correction algorithm is:First we use the feature vector extracted from the target image compared with the feature vectors of the image sample library to the select the most similar image in the database. Then we execute the anti-gamma correction on the original image with the gamma value of the image in the database to achieve the purpose of eliminating illumination effects. Preprocessing the uneven illumination video sequences with this algorithm can change the illumination effects of a video sequence well and reduce the influence of illumination.We use Mean Shift tracking algorithm for real-time tracking of the face regions in the first frame. Traditional Mean Shift algorithm is a non-parameter density estimation algorithm based on the gradient of the image. The algorithm uses the color histogram of the target region and the target candidate region for modeling after determining the target region in the first frame, and measures the similarity of the target model and the target candidate model with Pap function, and then find the best candidate for the target region iteratively. However, due to unevenness of illumination, the color histogram may fail to represent the characteristics of the object after the pretreatment of the video sequence. Based on the classic Mean Shift algorithm, we proposed a feature fusion based face tracking algorithm by combining the color histogram and edge direction histogram features.Finally, the tracking performance of the algorithm is compared with the classical Mean Shift algorithm. two sets of uneven illumination video were used in the experiments. The experiments’ result show that the proposed algorithm out performs the classical Mean Shift algorithm.