Clustering based detection of small target pedestrians for smart cities

Abstract

Detecting small targets in large images is challenging, especially for pedestrian detection in the field of self-driving cars for smart cities, for at least two reasons: (1) the pixels of small targets are blurry and are affected by the background; (2) the distribution of small targets is generally sparse and uneven, leading to the reduction of the detection efficiency. In this manuscript, we propose a clustering based scheme for the detection of small target pedestrians to address above difficulties. Clustering detection yields object cluster regions and gives an estimation of the object scales for these regions, so the clustered region is consistent with the input size of the fine detection model. Then, the detection results of small target pedestrians in the clustered region are obtained and restored to the input image to achieve the final detection results. Our detection strategy has several advantages over previous solutions. It saves a lot of computational power and running time which is vital for self-driving cars. Furthermore, the detection model is replaceable, making our scheme an easy solution to more complex tasks. We demonstrate the effectiveness of our scheme on both YOLOv4 and Faster R-CNN as their detection performance can be greatly improved.