An Object Detection Framework with Modified YOLOv3 for Camera Based Perception Systems

Abstract

An object detection framework based on modified YOLOv3 architecture is proposed for perception systems. In order to test the proposed object detector, a novel dataset consisting of road images for Indian road scenarios has been developed with various environmental conditions. The seven classes are considered in the object detection, in which five are vehicles frequently encountered during the Indian driving scenarios along with pedestrians and riders. The modified YOLOv3 promises the mean average precision of 84% for the detection of specified classes in the dataset. The F1 score and average IoU for the modified YOLOv3 are 81% and 71.78% respectively. The detection kernel is modified to train on the proposed dataset. The object detection results are compared with Faster Region-based Convolutional neural network (R-CNN), cascade R-CNN, single sort detector (SSD), baseline YOLOv3, and modified YOLOv3-tiny. The proposed object detector based on modified YOLOv3 yields improved mAP and hence it is more suitable for camera-based perception systems used in autonomous driving.