Agent-based framework for person re-identification

In computer based human object re-identification, a detected human is recognised to a level sufficient to re-identify a tracked person in either a different camera capturing the same individual, often at a different angle, or the same camera at a different time and/or the person approaching the camera at a different angle. Instead of relying on face recognition technology such systems study the clothing of the individuals being monitored and/or objects being carried to establish correspondence and hence re-identify the human object. Unfortunately present human-object re-identification systems consider the entire human object as one connected region in making the decisions about similarity of two objects being matched. This assumption has a major drawback in that when a person is partially occluded, a part of the occluding foreground will be picked up and used in matching. Our research revealed that when a human observer carries out a manual human-object re-identification task, the attention is often taken over by some parts of the human figure/body, more than the others, e.g. face, brightly colour shirt, presence of texture patterns in clothing etc., and occluding parts are ignored. In this thesis, a novel multi-agent based framework is proposed for the design of a human object re-identification system. Initially a HOG based feature extraction is used in a SVM based classification of a human object as a human of a full-body or of half body nature. Subsequently the relative visual significance of the top and the bottom parts of the human, in re-identification is quantified by the analysis of Gray Level Co-occurrence based texture features and colour histograms obtained in the HSV colour space. Accordingly different weights are assigned to the top and bottom of the human body using a novel probabilistic approach. The weights are then used to modify the Hybrid Spatiogram and Covariance Descriptor (HSCD) feature based re-identification algorithm adopted. A significant novelty of the human object re-identification systems proposed in this thesis is the agent based design procedure adopted that separates the use of computer vision algorithms for feature extraction, comparison etc., from the decision making process of re-identification. Multiple agents are assigned to execute different algorithmic tasks and the agents communicate to make the required logical decisions. Detailed experimental results are provided to prove that the proposed multi agent based framework for human object re-identification performs significantly better than the state of-the-art algorithms. Further it is shown that the design flexibilities and scalabilities of the proposed system allows it to be effectively utilised in more complex computer vision based video analytic/forensic tasks often conducted within distributed, multi-camera systems.