Real Time Adaptive Tracking System Using Computer Vision

Abstract

This project studies long-time object tracking in a sequence of frames. In this project, a detector is trained with specimens found on the path of a tracker that itself does not rely on the object detector. We attain high robustness and outdo current adaptive tracking-by-detection (11) approaches by decoupling object tracking and object detection. A substantial reduction of calculating time is attained by means of simple features for object detection and by using a cascaded method. The object location is marked in each frame. The task is to find the position of object in that frame else it must notify that the object is absent in the consecutive frames. We have developed a Real Time Tracking framework. The task of long-time tracking is divided as follows: Tracking, Learning and Detection. The tracker must follow the marked object of interest in consecutive frames. The detector restricts all observed appearances and amends the tracker when required. To evade these blunders there forth, the learning approximates the detector’s blunders and re-evaluates it. This project studies methods to recognize the detector’s faults and learn from, by developing a learning method with the help of “experts” which will estimate these blunders. We call it the P-N learning. With the help of RAT and P-N learning, our real-time processing can be described as an extremely integrated arrangement providing very precise object detection with RGB-D sensor.