Real time multi-object tracking using TLD framework

Today's state-of-the-art methods for object tracking perform adaptive tracking by detection, meaning that a detector estimates the position of an object and adjusts its parameters to the objects appearance at the same time. We propose a novel learning framework for tracking multiple unknown objects in a video stream by detection. Proposed system tracks multiple objects in presence of occlusion, clutter and scaling. The object is defined by its location and extent in a single frame. The tracker follows the object from frame to frame. The detector localizes all appearances that have been seen so far and corrects the tracker if required. The learning estimates detectors errors and updates that in the future to avoid these errors. A novel learning method (P-N learning) which estimates the errors by a pair of experts: (i) P-expert observes missed detections, and (ii) N-expert observes false alarms. First, instead of heuristically defining a tracking algorithm, we discovered that a discriminative structure prediction model from labeled video data and capture the interdependence of multiple influence factors. In every next frame the aim is to calculate the location and extent of object or indicate that object is not present. There are different algorithms which perceive the object in real-time. This system proposes a model which uses template matching algorithm with modifications based on SURF algorithm and squared difference error method. The template matching is done based on comparison of image features. We develop a novel method of tracking based upon template tracking algorithm which crops the region of interest(ROI) from the selected live object from a video stream from trained object database. Matching feature is found by applying principle component analysis.