Pedestrian Detection Using YOLOv5 For Autonomous Driving Applications

Abstract

Object detection is a branch of computer vision that permits us to detect and classify objects inside image or video. Pedestrian detection is an important segment of object detection, which is one of the the trending issues of computer vision and self-driving cars. Deep learning techniques furnished significantly enhanced results in the area of pedestrian detection. However, most of the literature reveals that the models used for addressing either speed or accuracy. In this paper, we addressed the speed and accuracy in pedestrian detection for autonomous cars. Manual inspection is replaced with a deep learning method, augmented images are exposed to You Look Only Once version 5 (YOLOv5) with activation function which creates a tradeoff between speed of detection and accuracy. This model is the best influential object detection algorithm at the moment to detect pedestrians in public places. The proposed pedestrian detection model is trained with the dataset of 'CityPersons' with 2975 images out of total 5000 images. The experimental analysis proves that the proposed algorithm remarkably improvises the detection speed with 0.011sec/image which can apply to the real time environment with a 46.2% Miss-Rate (MR) on highly occluded city persons dataset among various occlusion levels of it.