A feasibility analysis of a one-shot object tracking algorithm executed on a general-purpose single board computer

[摘要] ENGLISH ABSTRACT: Algorithms that are able to track generic objects in real-time have many useful applications such as security and traffic surveillance, augmented reality and sports analytics. Practical implications oftracking algorithms are further enhanced when the algorithms are able to be processed in real-time on mobile devices. Mobile SoCs are compact and energy efficient by design (Carroll, 2010) and present a possible implementation platform.Modern object tracking algorithms (Bertinetto, 2016) rely on computationally intensive convolutional neural network (CNN) architectures. CNNs are currently not able to be processed inreal-time on mobile devices (Lu, 2017). The research conducted in this thesis aimed to address the prior shortcoming in the computation of object tracking algorithms on mobile devices. A classically-designed object tracking algorithm, CMT, was chosen for investigation due to its flexibility inconfiguration of image features. CMT is independent of the method used to compute classical image features, permitting the usage of binary descriptor vectors that can be effectively computed. The primary investigation was the algorithm's suitability for implementation on a general-purpose heterogeneous computing platform. This was performed since heterogeneous platforms are common in mobile devices such as smartphones (Ignatov et al, 2016). A mobile platform was chosenbased on available hardware acceleration support and heterogeneous computing capacity.Baseline performance of 2.22 FPS was initially established on the chosen mobile hardware platform utilizing a strictly CPU execution model. An investigation into the optimal choice of image featuresrealized a 742% increase in FPS. The FPS was further increased through the utilization of on-board SIMD processors and achieved a real-time performance of 21.39 FPS. Due to OpenCV not supporting mobile GPU architecture, heterogeneous CPU-GPU acceleration on the mobile platform could not be investigated. When a desktop heterogeneous platform was utilized, the FPS throughput increased by 205% through heterogeneous CPU-GPU acceleration when compared to a CPU implementation. Results from an investigation into concurrent execution on the desktop platform did not meet theoretical expectations since the set of asynchronous GPU functions utilized did not execute completely asynchronously from the CPU.Real-time computation was achieved by utilizing strictly CPU execution on the mobile platform. The results of heterogeneous CPU-GPU acceleration on the desktop platform are transferrable to amobile platform, provided that the image processing library supports the mobile platform's heterogeneous capabilities. Thus, mobile devices are feasible platforms for real-time computation of classical object tracking algorithms due to the attained FPS, with further increases in FPS possible through heterogeneous CPU-GPU acceleration. This realizable increase in FPS through CPU-GPU acceleration indicates more computationally demanding algorithms can achieve real-time computation. Theoretical concurrent acceleration techniques were deemed to be of value as they present the upper limit achievable in a CPU-GPU heterogeneous execution model.

[发布机构] Stellenbosch University