Abstract

Successful transit planning in smart cities requires automated and efficient passenger counts at bus stops while also respecting the privacy of passengers—a paramount consideration in the age of responsible AI. In this thesis, we describe the implementation and development of a real-time passenger counting system at bus stops on Nvidia Edge devices powered only by solar panels with limited memory while not compromising privacy or incurring substantial costs. Numerous studies have developed and applied computer vision people detection techniques, although this has not been applied and optimized explicitly to edge-computing passenger counting at bus stops. In this regard, we evaluated different object detection models using a novel dataset from an overhead fisheye lens camera of passengers at a bus stop that we developed to analyze and improve the accuracy of the passenger counting system. We also optimize and reduce the models to allow them to be deployed on edge devices. We find that YOLO-V4 with mAP of 87% outperforms DetectNet-V2 and Faster-RCNN. The best object detection model has then been optimized and deployed on the Nvidia Jetson device, and the performance and efficiency of the passenger counting system have been evaluated. Deployment via Nvidia DeepStream on the edge showcased a more than 50% reduction in GPU, CPU, and memory consumption, enhancing efficiency while conserving energy. As a result, we present a more accurate and efficient edge-computing video analytics solution for an ethically responsible passenger counting system at the smart city bus stop.