Hierarchical Charged Particle Filter for Multiple Target Tracking

Multiple Object Tracking is a significant application of vision-based autonomous solutions. Although it is a logical extension of single object tracking, those algorithms do not simply scale over to the multiple object case. Tracking multiple objects brings on its own set of challenges and complexity.;A multiple object tracking system requires a set of tasks to be completed: (a) Objects need to be detected using Detection algorithms; (b) Objects need to be tracked using Tracking algorithms under conditions of object movement, occlusion, scale changes, illumination changes, scene movement etc.; (c) Such a system must perform as efficiently as possible, with the goal of real time execution. Combining these tasks presents its own set of challenges. This thesis provides a vision based system to track multiple objects under such variety of constraints.;A typical vision-based detection algorithm uses non-uniform convolution as one of its operations. This thesis provides a unique method to perform very fast convolution. The method called Stacked Integral Image builds upon the concept of box filters and integral image to accelerate convolution performed in vision based systems.;As tracking algorithms must handle changing scale/size of objects while tracking them, objects can become so small/blurred that geometry based detectors which use corners/edges etc. become unreliable. To utilize radiometric features for object detection, this thesis provides a unique method called Pattern Recognition by Cluster Accumulation that uses clustering and Hough style accumulation to reliably detect objects when they are small/blurred.;The particle filter is a powerful tool to track an object undergoing non-linear state dynamics with non-Gaussian noise. When tracking multiple objects using multiple particle filters, the particles of non-dominant targets tend to get hijacked by dominant targets. This problem is solved by a new resampling method called Charged Resampling which uses an electric charge like potential in a probabilistic setting to minimize particle hijack.;To handle moving objects in a moving scene, the problem of double dynamics is solved by employing a layered method called Hierarchical Particle Filter. This method cleanly separates scene tracking from multiple object tracking with a feedback connection to transfer intelligence from one sub-system to another.;Hence the novel contributions of this thesis are: (a) Fast convolution method, (b) Radio-metric detection algorithm, (c) Charged particle filter resampling, and (d) Hierarchical particle filter setup.