The Promise of Deep Learning for Indian Roads: A Comparative Evaluation of Architectures

Abstract

With the recent advances in the move towards autonomous vehicles, detecting vehicles through computer vision techniques is set to be an area of paramount importance. While existing studies apply artificial intelligence techniques for detecting vehicles in foreign road scenarios, most studies do not leverage the representational power of deep learning for Indian roads, wherein, the situation is often more challenging and complex. In this paper, we utilize an Indian road dataset to assess various deep learning models, focusing specifically on the condition of Indian roads at present. Alongside common vehicle classes like car, bus, motorcycle, etc., our dataset additionally consists of autorickshaws, trucks, person, and cycle, which are highly prevalent on Indian roads. Two detection algorithms: Faster RCNN and Single Shot Detector (SSD) have been incorporated with convolution neural net classifiers MobileNet, ResNet, and Inception Net, forming four detection models—Faster RCNN ResNet101, Faster RCNN ResNet50, SSD Inception NetV2, and SSD MobileNetV1. A comparative analysis of the algorithms shows that Faster RCNN ResNet 101 is the best performing model, achieving an accuracy of up to 57.60%, while SSD MobilenetV1 has the fastest inference time of 2.44 s. Our analysis shows that deep learning indeed has a bright future in the perspective of the Indian road