Comparing CNNs for non-conventional traffic participants

Abstract

This paper investigates performance of three state-of-the-art pretrained Convolutional Neural Network (CNN) models in terms of accuracy and latency for on and off-road obstacle detection in context of autonomous vehicle in Indian road. We investigated performance of Mask R-CNN, RetinaNet, and YOLOv3 on publicly available Indian road dataset. We evaluated accuracy and latency of these models on novel classes of objects such as animals, autorickshaws, caravan. Our results show that accuracy of Mask R-CNN is significantly higher than YOLOv3 and RetinaNet. We have also found Yolov3 is significantly higher than RetinaNet. We have also tested latency of the CNN models and found that latency of YOLOv3 is significantly lower than other two models and RetinaNet is significantly faster than Mask R-CNN. Finally, we have proposed an expert system to integrate environment parameters inside car along with outside car obstacles detected by YOLOv3 to estimate cognitive load of co-passengers of autonomous vehicle.