VRU-Net: Convolutional Neural Networks-Based Detection of Vulnerable Road Users

Research work on object detection for transportation systems have made considerable progress owing to the effectiveness of deep convolutional neural networks. While much attention has been given to object detection for automated vehicles (AVs), the problem of detecting them at road intersections has been underexplored. Specifically, most research work in this area have, to some extent, ignored vulnerable road users (VRUs) such as persons using wheelchairs, mobility scooters, or strollers. In this work, we seek to fill the gap by proposing VRU-Net, a CNN-based model designed to detect VRUs at road intersections. VRU-Net first learns to predict a VRUMask representing grid-cells in an input image that are highly probable of containing VRUs of interest. Based on the predicted VRUMask, regions/cells of interest are extracted from the image/feature maps and fed into the further layers for classification. In this way, we greatly reduce the number of regions to process when compared to popular object detection works such as Faster RCNN and the likes, which consider anchor points and boxes all over the image. The proposed model achieves a speedup of 4.55× and 13.2% higher mAP when compared to the Faster RCNN. Our method also achieves 9% higher mAP, comparing to SSD (Single Shot Multibox Detection).