Implementation of Compressed YOLOv3-tiny on FPGA-SoC

This paper presents the implementation of YOLOv3-tiny, a lightweight object detection algorithm on an FPGA-SoC embedded platform for real-time detection. First, YOLOv3-tiny is compressed by a model quantization method to reduce the memory size for speed improvement. Then, the backbone architecture, which is a heavy load process, is accelerated on the FPGA by applying the parallel pipeline method. The detection algorithm on the FPGA is evaluated with the performance metrics of the memory size, inference speed and accuracy. Inference on the self-collected pedestrian signals dataset showed that the proposed system could compress the memory size by 75% for real-time detection of 104.17 FPS.