| dc.description.abstract | Human following robots have been a very popular application, and with the recent popularity of deep learning and the development of the required hardware devices, deep learning based tracking algorithms are widely used in following robots applications. More and more tracking methods based on image processing have been proposed, which also include many deep learning methods that rely most on powerful computing resources such as GPU servers.
This paper focuses on the complexity of human tracking task by which a more efficient human tracking method is proposed, combining single-object tracker KCF, human detection module YOLO v3 and similarity comparison module to overcome the conflict of computational speed and accuracy in tracking task. In order to preserve the feasibility of flexible system changes in response to neural network development, we chose to use the HW/SW co-design based on Zynq SoC, with the PL (Programming Logic) part using AXI bus protocol to communicate with the PS (Processing System) part, and the PS part handling non-neural network computations and data transfers. The PL part deal with all neural network related computations. In addition, we introduced a new AI accelerator framework, Vitis-AI, and its Deep Processing Unit (DPU) in Zynq UltraScale + MPSoC ZCU104 to accelerate the YOLO v3 human detection module in the system. Finally, our human tracking approach can run at 11.5 FPS, achieving a 1.27x acceleration in system processing speed with the addition of a single-object tracker. Compared to the system on the CPU Intel Core i7700k@4.2GHz, the YOLO v3 human detection module on the ZCU104 accelerates 1.53 times faster while saving 87.1% in power consumption, reaching 409 GOPs on the ZCU104 and consuming only 15.57W, achieving a performance of 0.29 GOPS/s/DSP. | en_US |