Perceptual Modelling of Unconstrained Road Traffic Scenarios with Deep Learning

In recent times, advanced driver assistance system (ADAS) and autonomous driving have received significant interest in the automotive community. In this regard, lane assistance system and vehicle detection are considered as core modules of ADAS. However, the drawback is that the conventional image processing and vision-based techniques are quite slow and computationally expensive. The proposed work circumvents this limitation with the practical use of deep learning (CNN) based detection architecture. This paper proposes the use of CNN inspired detection methods such as Faster RCNN and YOLO for visual perception of road traffic. Notably, they were applied for effective vehicle detection in non-disciplined, heterogeneous Indian road traffic. This involved collecting own dataset for Indian urban heterogeneous traffic in both day and night time. An application of YOLO with VGG network resulted in a mAP score of 78.57%. On the other hand, Faster RCNN with Inception v2 and ResNet networks resulted in mAP score of 88% and 89.44%, on Indian road traffic datasets. This is a significant result; that shows the use of Deep Learning techniques for an efficient visual modelling of unconstrained road traffic scenarios.