Abstract

During the past decade, object detection and object tracking in videos have received a great deal of attention from the research community in view of their many applications, such as human activity recognition, human computer interaction, crowd scene analysis, video surveillance, sports video analysis, autonomous vehicle navigation, driver assistance systems, and traffic management. Object detection and object tracking face a number of challenges such as variation in scale, appearance, view of the objects, as well as occlusion, and changes in illumination and environmental conditions. Object tracking has some other challenges such as similar appearance among multiple targets and long-term occlusion, which may cause failure in tracking. Detection-based tracking techniques use an object detector for guiding the tracking process. However, existing object detectors usually suffer from detection errors, which may mislead the trackers, if used for tracking. Thus, improving the performance of the existing detection schemes will consequently enhance the performance of detection-based trackers. The objective of this research is two fold: (a) to investigate the use of 2D discrete Fourier and cosine transforms for vehicle detection, and (b) to develop a detection-based online multi-object tracking technique.

The first part of the thesis deals with the use of 2D discrete Fourier and cosine transforms for vehicle detection. For this purpose, we introduce the transform-domain two-dimensional histogram of oriented gradients (TD2DHOG) features, as a truncated version of 2DHOG in the 2DDFT or 2DDCT domain. It is shown that these TD2DHOG features obtained from an image at the original resolution and a downsampled version from the same image are approximately the same within a multiplicative factor. This property is then utilized in developing a scheme for the detection of vehicles of various resolutions using a single classifier rather than multiple resolution-specific classifiers. Extensive experiments are conducted, which show that the use of the single classifier in the proposed detection scheme reduces drastically the training and storage cost over the use of a classifier pyramid, yet providing a detection accuracy similar to that obtained using TD2DHOG features with a classifier pyramid. Furthermore, the proposed method provides a detection accuracy that is similar or even better than that provided by the state-of-the-art techniques.

In the second part of the thesis, a robust collaborative model, which enhances the interaction between a pre-trained object detector and a number of particle filter-based single-object online trackers, is proposed. The proposed scheme is based on associating a detection with a tracker for each frame. For each tracker, a motion model that incorporates the associated detections with the object dynamics, and a likelihood function that provides different weights for the propagated particles and the newly created ones from the associated detections are introduced, with a view to reduce the effect of detection errors on the tracking process. Finally, a new image sample selection scheme is introduced in order to update the appearance model of a given tracker. Experimental results show the effectiveness of the proposed scheme in enhancing the multi-object tracking performance.