A Voila-Jones face detector hardware implementation
This project attempts to realize a face detector using Voila-Jones algorithm. The reference C model is borrowed from 5kk73 GPU Assignment 2012, with some modify to fit hardware implementation and fixed some bug.
The code is written by Verilog/SystemVerilog and Synthesized on Xilinx KintexUltrascale FPGA using Vivado.
This code is just experimental for function, a lot of optimization can further be done.
Architecture
This project include 4 main module:
- The capture, this module captures the video input to frame buffer, with a scale of the input video
- The vj-fectch, this module accesses the frame buffer to get the pixels and feed into Volia-Jones engine.
- The vj, this is Volia-Jones engine, accepts original pixels and generates integral image and does classify
- The draw, this module is to draw a box onto face detected area. This draw engine does not operate on the frame buffer. It draws on video port on the fly.
Demo
Xilinx KCU105 Board, with HDMI input and output daughter boards.
The input 1080P HDMI input is scaled down to 480x270 to fill the frame buffer, then perform the face dection.
The final detection result is not very good. there is a lot missed and false detection.
Simulation
VCS is used
goto the sim/face subdirectory
run ./face
Synthesis
go to vivado/face directory
run "vivado -mode tcl -source face-ku.tcl"
TODO
- The area is large mainly because the ii registers. which can be optimized and the register width can be manual adjusted. Now all the registers use the same width.
- The detection speed is slow, more classifiers can be paralleled to speed up the performance.
- To speed up the performance, the data width of frame buffer should be enlarge to access multiple pixels in one clock.
- The II select MUX can be optimized.
- the Haar feature calculation can be implemented by shift and adder, not multiplier.
- The square root cal module is also a limit of speed.
Reference
https://sites.google.com/site/5kk73gpu2012/assignment/viola-jones-face-detection
Junguk Cho, "FPGA-Based Face Detection System Using Haar Classifiers"
Braiden Brousseau, "An Energy-Efficient, Fast FPGA Hardware
Architecture for OpenCV-Compatible Object
Detection"
Peter Irgens, "An efficient and cost effective FPGA based implementation
of the Viola-Jones face detection algorithm"
Author
LulinChen
lulinchen@aliyun.com
