VISUAL OBJECT TRACKING BASED ON NORMALISED CROSS CORRELATION AND LEAST SQUARE ESTIMATION

Abstract

In the discipline of computer vision, visual object tracking has been a hot topic for some time due to various applications such as observational analysis, athletics analysis, machine navigation, driverless vehicle, human-machine interfacing, and medical imaging etc. Various algorithms have been developed for visual object tracking in the direction of attaining good accuracy, speed and various other challenges that are faced in the implementation of good tracker like object shape, occlusion, changing scale etc. For tracking objects of interest particular detectors are used to detect objects and then tracked accordingly. Tracking process is generally completed by first Detecting and then Tracking. Observational analysis of objects in changing environments is the need of the hour. Tracked objects can also be classified while being tracked other than tracking only. Object tracking is achieved by observing objects’ spatial and temporal changes in a series of video frames along with its position and dimensions in frame. In observational analysis, object identification in sequence of frames is significant for tracing and analysing their behaviour. Moving object detection in a series of video frames is of prime importance in and tracking and background subtraction is a basic method for foreground separation. Various problems are faced in object tracking like illumination variation, object deformation, scale change, occlusion, background clutter, out of view, rotation etc. Although with the advent of deep learning technology and availability of high computational power in general purpose and high end embedded computers with very high graphical processing units, it is becoming feasible and more practical to design application of computer vision requiring object tracking which are faster and free from various problems faced in traditional methods at the cost of high computational requirements in training Deep Learning networks but the scarcity of computational power and faster response time in low end machines has led to research on this topic for an optimized algorithm that satisfies our requirement of faster and sufficient accurate tracker in the low end machines. To achieve this task, we have proposed a very simple and elegant fusion of cross-correlation with linear regression for finding the spatio- temporal location of the object in the series of frames. This fusion is cheap in computational cost which makes this algorithm faster compared to the existing ones at a bit of accuracy cost.