K-means based multiple objects tracking with occlusion reasoning

Abstract

This dissertation presents a novel and efficient multiple objects tracking technique dealing long-term and complete occlusion. The technique is based on a new low cost object appearance model to associate objects. This work is primarily focused on the improvement of resource utilization aspects, targeting real-time embedded applications with limited resources. Moreover, a comparative accuracy analysis is also performed to ensure that the proposed work is in close agreement with state of the art methods in terms of accuracy. This dissertation is mainly composed of three contributions. The first contribution presents a K-means based model for object appearance modeling. The appearance model combined with the simple object spatial position is used to infer the objects association/tracking decision. The objects appearance model and temporal spatial position is updated for one to one object/blob association throughout the sequence of video. However, for one-to-many blob/objects associations after the onset of occlusion a statistical distance measure is introduced for object association to deal occlusion. The comparative evaluation of resource utilization on standard datasets shows the superior performance of our approach in terms of computational time and/or memory as compared to the state-of-the-art baseline methods. In our second contribution, a low cost K-means based object appearance model is presented to achieve a faster solution for multiple objects tracking. In our first contribution and the recent literature, the object appearance is updated in every frame. However, there is no need to update the appearance model in every frame as more often it tends not to change. For this purpose, a novel histogram based appearance update model is applied on every detected blob to decide when to update the appearance model. Moreover, the employed histogram based cluster initialization further reduces the overall computational cost as the standard K-means algorithm can take more time to converge due to improper initialization of cluster centroids. The presented low cost model achieves much faster solution as compared to the baseline K-means and GMM based models with comparable memory requirements and accuracy. xi The first two contributions are about realizing a low cost multiple objects tracking method using merge and split approach for occlusion reasoning. The accuracy of the merge and split is compromised during occlusion. However, the emphasis of third contribution is on the improvement of accuracy during occlusion with reasonable resource utilization. In this contribution the K-means based method is extended to straight through approach, which tracks the individual objects despite occlusion thus increasing the accuracy. Low cost shape and appearance models are combined for pixel association during occlusion. Furthermore, two-pass outlier rejection technique is employed to address the issue of outliers. Our approach provides superior accuracy as compared to the state-of-the-art online multiple objects tracking approaches with comparable resource utilization. The prime objective of all the contributions of this dissertation is to provide a resource efficient solution of multiple objects tracking with comparable accuracy to the state of the art. In literature, the main focus of researchers is towards the aim of the accuracy with compromise on the resource utilization aspects of the system. The trending embedded smart cameras have started to replace the conventional PC based surveillance systems due to their low cost and built-in intelligence. To achieve real time solution with limited available resource, smart cameras require efficient algorithms with minimum resource utilization. With efficient resource utilization, the multiple objects tracking algorithms proposed in this dissertation, provides very strong grounds for the development of smart camera based real time surveillance solution.