An Embedded Deep Learning Based Traffic Advisory System

This paper presents a deep learning-based traffic advisory system relying on the count of vehicles on road at a certain time as the parameter advising people to take alternative routes as per requirement. There are some techniques that have led to a better inference time for a deep learning model but they are computationally expensive. Although we can afford to carry out the expensive computation on the cloud, this could hamper the performance of the real time traffic advisory system. In the proposed method implementation of two different deep learning frameworks - You only look once (YOLOv3) and Tiny YOLOv3 - to clock a quicker inference time while maintaining a significant level of accuracy and scalability of the system. Towards the end, we have presented our detection results on Indian driving dataset for vehicle detection. A comparative analysis of 4 deep learning techniques namely YoloV5, Ssd, Faster RCNN and EfficientDet has been performed in terms of performance.