koodg123 / hwac_object_tracker

FPGA accelerated TinyYOLO v2 object detection neural network

hwac_object_tracker

FPGA accelerated TinyYOLO v2 object detection neural network, capable of detecting 95 object classes. The design obtained the 5th place out of 65 teams, in the FPGA category, in the System Design Contest in Design Automation Conference 2018, San Fransisco (https://dac.com/content/2018-system-design-contest).

The final rankings are published in http://www.cse.cuhk.edu.hk/~byu/2018-DAC-HDC/ranking.html#final

The team list is in http://www.cse.cuhk.edu.hk/~byu/2018-DAC-HDC/teams.html

The design was deployed in the Xilinx PYNQ-Z1 platform (http://www.pynq.io/)

Design

The design is based on the TinyYOLO v2 Object Detection Neural Network (https://pjreddie.com/darknet/yolo/). We used Half-Precision Floating point (16 bit) our design. The implementation was done on Verilog HDL and using the Vivado 2017.2

The block design of our architecture is as follows,

The Vivado block design connecting our IP to the Zynq Processing System is as follows,

Resource Utilization

Resource Utilization :

Power estimate :

Repo Organization

• Images : contains the test images, annotations
• Others : contains documentation related files
• Results : contains the detection results
• hw : contains the RTL source files and the vivado projects
  • YOLO - contains the RTL sources and the Vivado project of TinyYOLO neural network implementation
  • TOP - contains the Vivado project with the top level block design
• py : contains the hardware overlay(.bit) and Jupyter Notebook, python libraries, executable on the ARM PS.