Performance Comparison of Different CNN models for Indian Road Dataset

Recent advancement in the field of computer vision and development of Deep Neural Network based object detection led researchers and industries to focus on autonomous vehicles. This paper aims to find how accurately previously proposed CNN architectures detect on-road obstacles in Indian road scenarios in the context of autonomous vehicle. We have compared three different convolution neural networks trained with COCO dataset for detecting autorickshaws in Indian road. We undertook statistical hypothesis testing to find effect of these three models, i.e. YOLOv3, Mask R-CNN, and RetinaNet on detection accuracy rate. While measuring accuracy, we have noted that detection accuracy rate of RetinaNet is significantly better than other two CNN architectures. Although there is no significant difference between other two networks in context of detection rate. The accuracy rate shows the performance of RetinaNet invariant to autorickshaws' color and shape, and different climatic and complex background scenarios.