An Improved Algorithm for Dense Object Detection Based on YOLO

Abstract

The YOLO v3 (you only look once) algorithm based on CNN (convolutional neural network) is currently the state-of-the-art algorithm that achieves the best performance in real-time object detection. However, this algorithm still has the problem of large detection errors in dense object scenes. This paper analyses the reason for the large error, and proposes an improved algorithm by optimizing confidence adjustment strategy for overlapping boxes and using dynamic overlap threshold setting. Experiments show that the improved algorithm has better performance in dense scenes while has little difference in other scenarios compared to the original algorithm.