An Approach for Object Tracking in Video Sequences

Abstract

In recent past there has been a significant increase in number of applications effectively utilizing digital videos because of less costly but superior devices. This
upsurge in video acquisition has led to huge augmentation of data, which are quite impossible to handle manually. Therefore, an automated means of processing these
videos is indispensable. In this thesis one such attempt has been made to track objects in videos. Object tracking comprises two closely related processes; object
detection followed by tracking of the detected objects. Algorithms on these two processes are proposed in this thesis.
Simple object detection algorithms compare a static background frame at pixel
level with the current frame in a video. Existing methods in this domain first try to detect objects and then remove shadows associated with them, which is a
two-stage process. The proposed approach combines both the stages into a single stage. Two different algorithms are proposed on object detection. First one to model
the background and the next to extract the objects and remove shadows from them. Initially, from first few frames the nature of each pixel is determined as stationary
or non-stationary and considering only the stationary pixels a background model is developed. Subsequently, a local thresholding technique is used to extract objects
and discard shadows.
After successfully detecting all the foreground objects, two different algorithms are proposed for tracking the objects and updating the background model. The
first algorithm suggests a centroid searching technique, where a centroid in current frame is estimated from the previous frame. Its accuracy is verified by comparing
the entropy of dual-tree complex wavelet coefficients in the bounding boxes of both the frames. If estimation becomes inaccurate, a dynamic window is utilized to
search for accurate centroid. The second algorithm updates the background using a randomized updating scheme.
Both stages of the proposed tracking model is simulated with various recorded videos. Simulation results are compared with the recent schemes to show the
superiority of the model.