Human Detection Methods In Still Images

Human detection has many applications in computer vision, such as smart cars, video surveillance and so on. An automatic method for finding humans in a scene serves as the first important preprocessing step in understanding human activity. The challenges are due to a wide range of poses that humans can adopt, large variations in clothing, as well as clustering background and complex illumination condition.Firstly, this paper proposes a human detection algorithm in still images. Take edge orientation feature and Haar-like feature as feature sets, and adopts an improved Adaboost algorithm to train strong classifier in each stage. Due to the slow training process of the conventional algorithm and the problem of cost sensitive, we propose a quick feature selecting algorithm to avoid calculating the classification error in the iteration, by means of establishing a table to store the information of feature. Combining with fisher discriminant analysis, we optimize the weights of weak learner, learn to acquire a new linear discriminant equation to maximize the separability of the different type of data and achieve the goal of optimizing the strong learner and decreasing the effect of cost sensitive. Experimental results show that the method speeds up the feature selection, and has good performance compared with the conventional Adaboost algorithm.Secondly, we improve the cascade structure of final classifier, by means of inserting additional stages into the original cascade structure. Support vector machine is used to train these new stages by exploiting the confidence of every strong classifier. Because of the full use of inter-stage information, we can exclude the negative detecting windows more quickly, and significantly enhance the efficiency of human detector.We apply the algorithm mentioned above to detect moving human in videos. The background subtraction based detection algorithm is applied. Gaussian Mixture Model is used to build the model of background. While updating the background model coefficient, we obtain the foreground image and the labeled moving region. After detecting moving human in the labeled region of original video image, we get the final results.Experimental result shows that the method proposed in this paper can detect human efficiently and accurately whether in still images or in video streams, and meet the need of real-time requirement.