Human Detection Based On Shape Of Gaussians Feature

Human detection has been one of the hottest research topics in computer vision areas recent years which mainly aim at making the computer find the pedestrians in images and videos automatically.As an important part of computer vision, human detection will be widely used in lots of areas, such as intelligent video surveillance, driver assistant system, content-based image retrieval, human-computer interaction, motion analysis and so on.Human detection based on statistical classification has been the development trend in recent years, which uses machine learning methods to train and learn automatically from a large set of features extracted from training samples for a classifier to distinguish between human and non-human, which is a pattern classification problem in fact. The key to human detection based on statistical classification is to choose the feature which has strong ability to distinguish and the appropriate machine learning methods. A human detection algorithms presented in this paper is also based on statistical classification, which takes the shape of gaussian as feature and uses Adaboost and improved Logitboost to train for a classifier of cascade structure to detect pedestrians. Because shape of gaussian feature is built based on covariance feature by adding factor of mean, the fact that shape of gaussian feature has better ability to distinguish than covariance feature is proved by theory and experiment. Further analysis shows that the shape of gaussian feature has the structure of a lie group topology, so we prefer measuring distance between the features based on lie group theory to measuring based on traditional vector space and an improved Logitboost based on the characteristics of lie group of the shape of gaussian feature is applied to train the features when we build the cascade classifier.In addition, in order to meet the requirements of real-time human detection, we parallel both the process of training for human classifier and the process of human detection with the TBB library in the implementation of the algorithm, except that building cascade classifier for fast classification and introducing the method of integral image to quickly extract the features.Experimental results show that the proposed human detection algorithm performs well, and also can detect pedestrians basically in real time.