A Survey of Dense Object Detection Methods Based on Deep Learning

Object detection based on images or videos is one of computer vision’s most common and practical tasks. In recent years, thanks to the development and advancement of deep learning methods, object detection algorithms and frameworks have made significant breakthroughs, allowing many detection methods to be applied to real-world scenarios. However, traditional object detection algorithms perform poorly in complex scenarios such as densely populated and crowded environments, which are frequently encountered in practical applications. Consequently, dense object detection has emerged as an important research branch within object detection, garnering increasing attention from scholars. New ideas and methods for different application scenarios and scientific problems continue to emerge. This paper comprehensively surveys the latest advancements in dense object detection, summarizing and analyzing the most advanced methods. We discuss the similarities and differences, as well as the advantages and disadvantages, of dense object detection algorithms based on five different approaches: NMS, LOSS, re-scoring, transform and YOLO. Additionally, we introduce several datasets and evaluation metrics relevant to dense object detection tasks. Finally, we summarize the current challenges in dense object detection research and discuss future research directions.