The Research On Color Blob Based Pedestrian Classification For Video Surveillance

Foreground classification in surveillance is always the hot topic in surveillance processing research area. Foreground classification is useful for video summarization, object tracking and behavior understanding. Video summarization makes short and precise description about the content of the videos, classify the foreground people in surveillance can help video summarization more precise, which can be viewed as the basement of video summarization. Object tracking aims to realize track the object under different conditions. It is carried on under the assumption that the tracked object is determined. Therefore, the classification of foregrounds, which simplify the process of object tracking, will make the object tracking more precise and more efficient without the interaction from the users. Behavior understanding distinguishes the meaning of pedestrian behavior by establishing behavior model after training on the database. Pedestrian classification can realize multi angles surveillance foreground classification which will discover the connections between similar pedestrians and improves the efficiency of behavior understanding.Foreground classifications center around how to classify different kinds of object preciously for a long time, it can distinguish the foreground into cars, vans, bicycles, a group of pedestrians or single pedestrian. All these researches adapt appearance based and velocity features, which will be very different as they are different categories, to represent the foreground and then build classification models after training on the databases, which can be used for foreground classification. However,these algorithms can only realize the foreground rough classification and they will fail when we want to do more detailed classification on foregrounds. Besides, existing works needs to do a lot of training before they classify the foreground, which will not be suitable for high efficient application.Our work can not only distinguish the pedestrians and non-pedestrians, also give more elaborate classification on pedestrians. We propose a method which will classify the pedestrian based on color blob modeling and subspace clustering by rank. Firstly, we apply color blob modeling on the persons so that we can model the foreground person, which can further help to set separator lines between the torso and the bottom. This will add spatial information to features and brings us higher precision. Secondly, we use canonical correlation to grab the global information of features, which can make the data feature more pairwise. In the end, we use subspace clustering algorithm based on rank to classify the persons, which proved very robust to noise and transferring in feature data. We compare our algorithm with other methods on several databases. Our method proves to be the most precise one.