A Lightweight Network for Detecting Pedestrians in Hazy Weather

Abstract

Most of the existing detection models fails to detect pedestrians in hazy weather conditions. Therefore, to improve the safety of the driver in semi-autonomous or autonomous vehicles, a lightweight network has been proposed which can detect pedestrians effectively in hazy weather. A lightweight network YOU-ONLY-LOOK-ONCE-v2 (YOLOv2) + MobileNetv2 + Convolutional Block Attention Module (CBAM) was proposed. To build more efficient and faster model, YOLOv2 is employed, and to reduce both number of parameters and computational complexity, we adopt MobileNetv2 as our backbone model. In the proposed model, bounding box loss error is optimized by applying normalization and we also introduced attention module (CBAM) to improve the detection accuracy. Prior to the model training, we applied K-means clustering algorithm to figure out optimal number of prior anchor boxes in our dataset. Experimental results show that the proposed network achieves 87.4% average precision and detection accuracy on hazy person dataset and runs with 173.6 frames per second (FPS).