YOLOv4 Pedestrian Target Detection Based on Embedded Platform

Due to the dense number of pedestrians in complex environments and the predominance of small moving targets, the YOLOv4 algorithm has low detection accuracy and the large number of model parameters cannot be deployed on mobile devices with limited resource computing power. In this paper, we propose a YOLOv4 pedestrian target detection method based on the embedded platform Jetson Nano, using the ONNX (Open Neural Network Exchange) and TensorRT algorithms to optimize the YOLOv4 network model for inference; The trained network model is deployed on the Jetson Nano platform to build a target detection system; Using an SSD (Solid State Drive) with SATA3 interface to add a Swap function, which can ameliorate the storage structure of the Jetson Nano platform. The finally experimental results show that YOLOv4 pedestrian detection based on the embedded platform Jetson Nano achieves 94.39% mAP on the VOC extraction dataset, and the inference time is shorter, the IO throughput of the system is improved, and the robustness is enhanced.