oa Visual Scale-adaptive Tracking For Smart Traffic Monitoring

This paper presents a novel real-time scale adaptive visual tracking framework and its use in smart traffic monitoring where the framework robustly detects and tracks vehicles from a stationary camera. Existing visual tracking methods often employ a semi-supervised appearance modeling where a set of samples are continuously extracted around the vehicle to train a discriminant classifier between the vehicles and the background. While proving their advantage, many issues are still to be addressed. One is a tradeoff between high adaptability (prone to drift) and preserving original vehicle appearance (susceptible to tracking loss with significant appearance variations). Another issue is vehicle scale changes due to perspective camera effect which increases the potential for an inaccurate update and subsequently visual drifting. Still, scale adaptability has received little attention in vision-based discriminant trackers. In this paper we propose a three-step Scale Adaptive Object Tracking (SAOT) framework that adapts to scale and appearance changes. The framework is divided into three phases: (1) vehicle localization using a diverse ensemble, (2) scale estimation, and (3) data association where detected and tracked vehicles are correlated. The first step computes vehicle position by using an ensemble based on a compressed low-dimensional feature subsets projected from high-dimension feature space by random projections. This provides the diversity needed to accommodate for individual classifiers errors and different adaptability rates. The scale estimation step, applied after vehicle localization, is computed based on matched points between a pre-stored template and the localized vehicle. This doesn't only estimate the new scale of the vehicle but also serves as a correction step to prevent drifting by estimating the displacements between correspondences. The data association step is subsequently applied to link detected vehicle of current frame with the tracked vehicles. Data association must consider factors like the absence of detected target, false detections and ambiguity. Figure 1 illustrates the framework in operation. While the vehicle detection phase is executed per frame, the continuous tracking procedure ensures that all vehicles in the scene, no matter how complex it is, are correctly accounted for. The performance of the proposed Scale Adaptive Object Tracking (SAOT) algorithm is further evaluated with a different set of sequences with scale and appearance changes, blurring, moving camera and illumination. SAOT was compared to three established trackers in the literature: Compressive Tracking , Incremental Learning for Visual Tracking and Random Project with Diverse Ensemble Tracker using standard visual tracking evaluation datasets [4]. The initial target position for all sequences was initialized using manual ground truth. Centre Location Error (CLE) and Recall are calculated to evaluate the methods. Table 1 presents the CLE errors and recall in parentheses measured on a set of 2 sequences with different challenges. It clearly demonstrates that SAOT performs better than the other trackers.