Improvement Of The Human Detection Algorithm Based On Histograms Of Oriented Gradients

Human detection is a technique that allows the machine to study where people exist in images or videos by the way of people thinking. As an important part of the human target recognition and tracking, it has broad application prospects in video surveillance, video identification, driver assistance systems, athlete motion analysis and other fields. But how quickly and accurately detect the human body to the researchers is not a small challenge, due to the variability of the body’s own particularity and the environment’s complexity where they are in. We can conclude that human detection based on HOG has more advantages than other methods which used to detect object from the result of Pascal VOC Challenge. However, the method also has the following disadvantages.1. It is easily influenced by the environmental factors such as illumination and weather conditions.2. It is difficult to satisfy the requirement of real-time due to the slowly processing speed.3. There is no better solution to solve the problem of shade. This article will focus on solving the two previous problems.A deep research made in view of the human detection algorithm which based on HOG, we put forward the following improvements on the front two shortcomings.1. Proposed human detection algorithm based on HOG-PCA and PCA-ICA in order to reduce the dimension of original algorithm and speed up the calculation. Both the algorithms employ the principal component analysis to reduce the dimension and use the SVM to classify the sample. In order to keep edge information as much as possible, the former adopt the DoG filter for filtering process to the final features.Finally, features which SFS algorithm choose are fed into SVM classifier. The algorithm based PCA-ICA directly reduce the dimension of HOG features. Extract the reduced features by independent component analysis as the input of SVM classifier.2. Proposed human detection algorithm which based on improved HOG features in order to lessen the influence of illumination on algorithm. First of all, the algorithm divided sample images into three types according to its gray average and standard deviation, including abundant light, weak light and normal light types. Two differentillumination compensation methods are proposed for the images which are divided into strong and weak light. Combining masks tested various 1-D point derivatives uncentred[-1, 1] with centered [-1, 0, 1] calculate gradient so as to highlight the edge features. A more detailed division will be done by a new way to blocks and cells which are used to count on gradient histograms.In order to verify the effectiveness of the algorithm which proposed in this article,we use the INRIA person dataset testing the validity of the algorithm. All experimental results are obtained under Matlab2012 a.