Person re-Identification over distributed spaces and time
Replicating the human visual system and cognitive abilities that the brain uses to process the information it receives is an area of substantial scientific interest. With the prevalence of video surveillance cameras a portion of this scientific drive has been into providing useful automated counterparts to human operators. A prominent task in visual surveillance is that of matching people between disjoint camera views, or re-identification. This allows operators to locate people of interest, to track people across cameras and can be used as a precursory step to multi-camera activity analysis. However, due to the contrasting conditions between camera views and their effects on the appearance of people re-identification is a non-trivial task. This thesis proposes solutions for reducing the visual ambiguity in observations of people between camera views This thesis first looks at a method for mitigating the effects on the appearance of people under differing lighting conditions between camera views. This thesis builds on work modelling inter-camera illumination based on known pairs of images. A Cumulative Brightness Transfer Function (CBTF) is proposed to estimate the mapping of colour brightness values based on limited training samples. Unlike previous methods that use a mean-based representation for a set of training samples, the cumulative nature of the CBTF retains colour information from underrepresented samples in the training set. Additionally, the bi-directionality of the mapping function is explored to try and maximise re-identification accuracy by ensuring samples are accurately mapped between cameras. Secondly, an extension is proposed to the CBTF framework that addresses the issue of changing lighting conditions within a single camera. As the CBTF requires manually labelled training samples it is limited to static lighting conditions and is less effective if the lighting changes. This Adaptive CBTF (A-CBTF) differs from previous approaches that either do not consider lighting change over time, or rely on camera transition time information to update. By utilising contextual information drawn from the background in each camera view, an estimation of the lighting change within a single camera can be made. This background lighting model allows the mapping of colour information back to the original training conditions and thus remove the need for 3 retraining. Thirdly, a novel reformulation of re-identification as a ranking problem is proposed. Previous methods use a score based on a direct distance measure of set features to form a correct/incorrect match result. Rather than offering an operator a single outcome, the ranking paradigm is to give the operator a ranked list of possible matches and allow them to make the final decision. By utilising a Support Vector Machine (SVM) ranking method, a weighting on the appearance features can be learned that capitalises on the fact that not all image features are equally important to re-identification. Additionally, an Ensemble-RankSVM is proposed to address scalability issues by separating the training samples into smaller subsets and boosting the trained models. Finally, the thesis looks at a practical application of the ranking paradigm in a real world application. The system encompasses both the re-identification stage and the precursory extraction and tracking stages to form an aid for CCTV operators. Segmentation and detection are combined to extract relevant information from the video, while several combinations of matching techniques are combined with temporal priors to form a more comprehensive overall matching criteria. The effectiveness of the proposed approaches is tested on datasets obtained from a variety of challenging environments including offices, apartment buildings, airports and outdoor public spaces.
AuthorsProsser, Bryan James
- Theses