Efficient Multi-Receptive Pooling for Object Detection on Drone

Abstract

Object detection is the most fundamental and important research in computer vision to discriminate the location and class of the object in the image. This technology has been continuously researched for the past few years. Recently, with the development of hardware such as GPU computing power and cameras, object detection technology is gradually improving. However, there are many difficulties in utilizing GPUs on low-cost devices such as drones. Therefore, efficient deep learning technology that can operate on low-cost devices is needed. In this paper, we propose a deep learning model to enable real-time object detection on a low-cost device. We experiment to reduce the amount of computation and improve speed by modifying the CSP Bottleneck and SPPF parts corresponding to the backbone of YOLOv5. The model has been trained on MS COCO and VisDrone datasets, and the mAP values are measured at 0.364mAP and 0.19mAP, which are about 0.07 and 0.04 higher than Refinedetlite and Refinedet, respectively. The speed is 23.010 frames per second on the CPU configuration, which is enough for real-time object detection.