An Efficient Algorithm for Real-Time Vehicle Detection Using Deep Neural Networks

Abstract

Vehicle detection is one of the major tasks in the field of Computer vision and Intelligent video surveillance systems. In this paper, we present a novel approach to detect vehicles and classify them. The proposed system is based on the YOLOv3 object detection algorithm. The backbone network we used for Feature Extraction is pretrained DenseNet121. The YOLOv3 network is further tweaked on the detection layer by pacing an extra prediction scale to form the desired architecture of our research. We also incorporated Distance IoU loss and DIoU-NMS in the network which further boosts the accuracy of the network. This network has four scales of prediction layers, with the fusion of these layers, networks predictive function increases. The proposed network is tested by modifying various parameters to find the optimal results.  Our network is proven to be effective in the detection of vehicles during night time, long-distance vehicles, and occluded vehicles. Our experiments on PASCAL VOC 2007, 2012 and COCO datasets achieves desired results in the improved network. We also created a dataset based on night-time vehicle images from Traffic CCTV footages to maximize detection accuracy during extreme weather conditions, especially at night. The Improved YOLOv3-Net model with input size 608 × 608 accomplishes high mAP of 82.8 % which is a promising result in detecting real time videos.