Design of embedded real-time human detection system based on Zynq platform

Abstract

Abstract With the continuous breakthroughs in intelligent algorithm models, the computational complexity of target detection algorithms has continued to increase. At the same time, the performance and power consumption requirements of edge systems for hardware devices have become increasingly stringent. To better weigh the detection speed and accuracy of the artificial intelligence algorithm deployed on the embedded Zynq platform, we streamline and optimize the YOLO-v4 network model and then retrain the model based on the COCO dataset. Finally, the quantification and compilation processes are carried out in sequence to realize model migration and deployment from the server side to the Zynq device. Through experimental verification, the detection time of the modified YOLO-v4 algorithm model in this paper is 31.42 ms, and the detection error is only 4.0 pixels.