UltraHiNet, LLC (UHN) proposes to develop a prototype next generation tracking architecture which exploits wide area motion imagery (WAMI) systems and leverages projected heterogeneous high performance computing (HPC) architectures for on-board and ground-based processing of vehicle and dismount targets for urban surveillance. Our research team will focus on the following goals: 1. Develop parallel GPU-based implementations of compute intensive video processing algorithms for dense object tracking that are close to the sensor for on-board processing of WAMI, and 2. Develop parallelized multitarget tracking using novel computational data association algorithms for producing reliable tracklets that can be further processed to long tracks in cluttered urban environments. Visual information using salient object features will enable persistent tracking through complex scenes with visual distractors, large object motion displacements, long occlusions, noisy detections, and shadows. Appearance-based matching and feature fusion methods will be investigated to track moving vehicles by fusing multiple appearance features such as intensity, edges, shape, texture, spatial and temporal context to resolve data association ambiguities within dense collections of movers. We will analyze parallel processing architectures for real-time computation/applications, to determine the fastest processor with best numerical quality, and the number of processors in a cluster necessary for real- or near-real time execution. Finally, we will address issues related to optimal utilization of limited on-board resources through proper load balancing, data driven regular recursive image tiling and adaptive program structures. Thus, we will effectively demonstrate the feasibility of our algorithm(s) in meeting Air Force Research Lab needs and provide a Phase II development plan with performance goals and key technical milestones.
Benefit: If successful, the proposed research will constitute a significant breakthrough in the solution of problems related to multiple hypothesis multitarget tracking by using modern parallel computing architectures. Our methodology will be optimized for execution on highly parallel multicore processors such as multicore CPUs, manycore graphics processing units (GPUs) and hybrid multicore processors (HMPs). We will also include detailed analysis (in Phase II) of numerical quality. This research and development applies to a wide variety of military and domestic applications, such as ship, submarine, torpedo, and towed array design and performance optimization, as well as domestic applications such as bridge pier design and monitoring in rivers or estuaries having high current flow. UltraHiNet LLC will license or sell the Phase II product to one or more aerospace firms or defense contractors.
Keywords: Multi-Target Tracking, Wide Area Motion Imagery, Processing Unit, Parallel Computing, High-Performance Computing
