Methods for Feature Selection and Appearance Model in Visual Object Tracking.

Abstract

The objective of visual object tracking is to find the location, orientation and scale (size) of an object in successive video frames. There are several applications of object tracking in computer vision. Some of them include human-computer interaction (HCI), visual surveillance, action recognition, vehicle navigation, video compression and face recognition.

The object tracking has been an active research area due to several challenges involved. One of the most encountered challenges is the object appearance variation which occurs from one frame to another in the form of deformation, motion blur, pose change, and illumination variation. Scale change, partial occlusion and clutter are the other challenges that an object tracking algorithm has to deal with.

Therefore, the appearance model is the most important component in the visual object tracking. Based on the learning mechanism used, a tracker can be called as generative or discriminative. The generative tracking method learns an object appearance model, composed of different object variations, without considering its surrounding background.The discriminative object tracking is based of construction of discriminative classifier using both target instance as well as the instances from the surrounding background. An object tracker can use both mechanisms for better tracking performance.
This research work focuses on discriminative feature selection and online appearance model construction for robust visual object tracking. A fast online algorithm to select discriminative feature in multiple instance learning (MIL) framework is proposed. The algorithm is based on maximizing the discriminative classifier score. The tracking performance is better than the existing features election methods in this framework. Unlike other methods, it does not use sigmoid function, and therefore, it is suitable for VLSI implementation. The use of kernel trick on Haar-like features for target tracking is introduced. Furthermore, use of Haar-features in half target spaces is explored. In the same MIL framework, to get the actual target size in successive video frames, a scaling strategy is applied. By having the feature matching using kernel, Haar-features in half target spaces and scale adaptation method in a single tracking framework, a better tracking performance is achieved, as compared to the state-of-the-art trackers.

Visual tracking algorithm using discriminative and generative appearances is explored. The discriminative model is a linear support vector machine(SVM) classifier trained in the first video frame only. In the successive frames, online appearance model is learned using similarity measure between newly tracked sample, negative examples and parameter of the trained SVM. Based on learned appearance model, a likelihood model is constructed for selecting a tracking instance. The appearance model is also learned on HOG features using discriminative parameter update.
A training algorithm is proposed to online learn the parameter vector of SVM when new examples are available in each frame of a video. This is an iterative method which is based on maximizing the sum of projection lengths of a positive example and a negative example which are closest to the hyperplane formed by parameter to be updated. Using learned parameter vector, likelihood model is constructed and applied to get the tracking instance. An object representation is also learned based on sparse DCT coefficients. The learned structure is to lower the effect of occlusion while preventing the basic appearance. Using sparse 2-dimensional discrete cosine transform (2D DCT) coefficients as discriminative features, a visual object tracker is proposed. A simple method to select the discriminative DCT features is also proposed. The discriminative coefficients are selected in each frame based on their mean of probabilities with parameters of positive (target) and negative (background) instances. A discriminative classifier is constructed by using selected features. Some intermediate tracking instances are obtained by (a) computing feature similarity using kernel and (b) finding the maximum classifier score. The final tracked instance is obtained by comparing their 1-dimensional array correlation with the raw pixel based appearance models. The final tracked instance is also obtained based on score of adiscriminative classifier learned in successive frame susing HOG features.
A method to train a weight vector is proposed. It is based on kernel similarity measure between parameter vector and examples. In each video frame, the parameter vector is updatedif(a)thereceivedpositiveexamplehaslowersimilarityscorethanthatofprevious positive example with lowest score and(b)received negative example has higher similarity score than that of previous negative example with highest score. The similarity scores of such example vectors (positive and negative) are also used to construct a discriminative model. Also, successive addition of tracked samples weighted with their similarity scores in each frame forms a raw pixel based appearance model. The proposed likelihood model for finding the target in next frame is composed of learned discriminative and pixel based appearance models. A modified parameter learning method seek to maximize the score as well as margin between the examples. The appearance model learned using HOG features gives better tracking performance.
The tracking accuracy of all the proposed methods is better than state-of-the-art trackers in a number of challenging video sequences.