YOLO-Log: A Light-weight Object Detector for Logistics Safe Driving

Object detection algorithms are attracting more and more attention in the application of intelligent logistics, but current object detection algorithms have problems such as high computational cost, low detection accuracy, and difficulty in deploying edge devices with limited computing resources in complicated logistics scenarios. This paper proposes a logistics object detection network (YOLO-Log) with a light-weight residual structure GhostNet. YOLO-Log is based on YOLOv5, which incorporates a light-weighted residual structure in Backbone and substantially reduces the model parameters of Backbone. Then, a simple and effective weighted bi-directional feature pyramid network is invoked to enhance the network to extract feature information of the target from the complicated logistics background. Finally, in light of the current lack of publicly available logistics object detection datasets, we collected and produced the logistics object detection dataset Logistics-3k. YOLO-Log achieves 91.4% mAP on the Logistics-3k dataset, which reduces the model parameters by 14.2% compared to YOLOv5s, and the detection time is only 10.2ms, addressing the industrial requirements of logistics scenarios.