University of Twente Student Theses


A Tunable Accelerator for the YOLOv4-tiny Object Detector using Vitis Unified Software Platform

Balamurali, Sridhar (2023) A Tunable Accelerator for the YOLOv4-tiny Object Detector using Vitis Unified Software Platform.

[img] PDF
Abstract:Deep learning-based object detection plays a crucial role in various computer vision applications. However, deploying these complex networks on embedded systems that have limited resources presents significant challenges. The accelerator is designed to optimize the performance of not only the computationally intensive convolutional layers but also other non-convolutional layers in the YOLOv4-tiny model using Vitis unified software platform. Key optimizations, such as fixed-point quantization and channel interleaving, are employed. The Vitis unified software platform is utilized to dynamically configure the layers of YOLOv4-tiny within the processing system of the ZedBoard. The proposed accelerator achieves a significant speedup, with the convolutional layers running 20 times faster compared to previous works on the same platform with a MACC operation taking 2 clock cycles inside the convolution block with a throughput of 5.84 GOPS/secs, resulting in an inference rate of 3.3 seconds per image. The overall architecture achieves a throughput of 2.05 GOPS/sec with resource utilization of 166 (76%) DSP units, 149 (53%) BRAM18K blocks, LUT, and FF utilization of about 56% and 43% respectively. Furthermore, when compared with the ARM A9 processor on the ZedBoard and host CPU implementation, the implemented architecture demonstrates a speed improvement of 58x, and 3x respectively.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Programme:Electrical Engineering MSc (60353)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page