Deep Learning Approaches on Pedestrian Detection in Hazy Weather

Effectively detecting pedestrians in various environments would significantly improve driving safety for autonomous vehicles. However, the degrpted visibility and blurred outline and appearance of pedestrian images captured during hazy weather strongly limit the effectiveness of current pedestrian detection methods. To solve this problem, this article presents three novel deep learning approaches based on you only look once. The depth wise separable convolution and linear bottleneck skills were used to reduce the computational cost and number of parameters, rendering our network more efficient. We also innovatively developed a weighted combination layer in one of the approaches by combining multiscale feature maps and a squeeze and excitation block. Collected pedestrian images in hazy weather were augmented using six strategies to enrich the database. Experimental results show that our proposed methods can effectively detect pedestrians in hazy weather, significantly outperforming state-of-the-art methods in both accuracy and speed.