AUTOMATED LONG-TERM OBJECT TRACKING WITH ONLINE LEARNING

Abstract

This thesis work looks into automatic detection as well as long-term tracking of any unknown object in a video sequence. Every object is described by position and area covered in any particular frame. Bounding box defines the object of interest in the first frame. For automatic detection of object, Gaussian-Mixture-Model for background subtraction is used, this makes the system suitable for use in automatic surveillance and monitoring. In consecutive frames, objective is to find objects position and area or to point out objects absence when not present. This tracking approach fragments the long-term tracking task into simpler subtasks of tracking-learning-detection. The tracker tracks object in every consecutive frame. Detector localizes every appearance that is observed so far and makes tracker error free by correcting when necessary. Neither tracking nor detection can single-handedly give solution to the long term tracking problem. Learning removes the detector’s errors and also updates the detector to overcome future errors. This work studies way to find detector’s error along with learning from it. The novel online learning approach (P-N learning) removes errors by pair of ‘experts’: a) P-expert finds out the missed detections b) N-experts finds out false alarms. The process of learning is semi-supervised learning with a set of labelled data and we need to label the unlabelled one. The TLD framework along with P-N learning is described. This method is different in the way that here the classifier is trained online, hence this method is suitable for tracking any unknown object. The outcome is real time tracking which enhances with time. This framework is advertised under Predator i.e., a smart camera that learns with time.