Fast Pedestrian Detection for Real-World Crowded Scenarios on Embedded GPU

Abstract

The behavior of individuals in crowds in public places has gained enormously in importance last year, for example through distancing requirements. However, automatically detecting pedestrians in real-world uncooperative scenarios remains a very challenging task. Especially crowded areas in surveillance footage are not only challenging for automatic vision systems, but also for human operators. Furthermore, complex detection models do not scale easily and are not traditionally designed for on-device processing in resource-constrained smart cameras, which become more and more popular due to technical and privacy issues at large events. In this work, we propose a new Fast Pedestrian Detector (FPD) based on RetinaNet which is a fast and efficient architecture for embedded platforms. The proposed FPD provides near real-time and real-time detection of hundreds of pedestrians on embedded platforms, outperforming popular YOLO-based approaches traditionally tuned for speed. Furthermore, by evaluating our approach on several different Jetson platforms in terms of speed and energy profiles, we highlight the challenges related to the deployment of a deep learning based pedestrian detector on embedded platforms for smart surveillance cameras.