E-scooters have become ubiquitous vehicles in major cities around the world. The numbers of e-scooters keep escalating, increasing their interactions with other cars on the road. Compared to traditional vulnerable road users, like pedestrians and cyclists, e-scooter riders not only have different appearances but also behave and move differently. This situation creates new challenges for vehicle active safety systems and automated driving functionalities.
Detection is the first step for intelligent systems and AI algorithms to mitigate the potential conflicts with e-scooter riders. In this project, we propose a small benchmark dataset for e-scooter rider detection task, and a trained model to support the detection of e-scooter riders from RGB images collected from natural road scenes. The dataset contains equal number of cropped images of e-scooter riders and other vulnerable road users. For the e-scooter rider detector, we propose an efficient pipeline built over two existing state-of-the-art convolutional neural networks (CNN), You Only Look Once (YOLOv3) and MobileNetV2. We fine-tune MobileNetV2 over our dataset and train the model to classify e-scooter riders and pedestrians. We obtain a recall of around 0.75 on our raw test sample to classify e-scooter riders with the whole pipeline. Moreover, the classification accuracy of trained MobileNetV2 on top of YOLOv3 is over 91%, with precision and recall over 0.9.