Vulnerable road users detection based on convolutional neural networks

The roads support many different types of users. With geared efforts to advance connected and autonomous vehicles (CAVs), smart mobility systems, and advanced driver assistance-based vehicles, the safety of road users becomes a growing concern. Some users may require more cautious interactions and support to ensure safe usage of the road infrastructures. While considerable effort has been done to detect different types of road users or objects from a vehicle's viewpoint, there are certain classes of vulnerable road users which have been overlooked in prior works. The objective of this work is to detect vulnerable road users (e.g., Strollers, Motorbikes, and Bicycles) in order to aid in reduction of collisions. We investigate the performance of one-stage and two-stage deep object detection methods in detection of said vulnerable users. Since there is a lack of publicly accessible datasets containing objects of our interest from an infrastructure viewpoint, we introduce our own dataset collected from a road side. We highlight the benefits and shortcomings of the studied methods in the context of vulnerable road users detection under challenging conditions such as occlusions.