Real-time Pedestrian Detection Using Resource Constrained Embedded Platforms – A Review

Autonomous Vehicles (AV) are currently limited to the sensors fitted on the vehicle itself to detect pedestrians instead of working collaboratively with the road side units (RSUs). If Resource Constrained Embedded Platforms (RCEPs) were capable of achieving real-time pedestrian detection, they could be deployed at busy traffic intersections to assist AVs. Currently, high-end GPU platforms are required to achieve real-time performance, but their size and cost mean they are not viable for deployment at busy traffic intersections. Therefore, this study performs a comparative analysis as to the ability of RCEPs to achieve real-time pedestrian detection using Convolutional Neural Network (CNN) architectures. Several CNN architectures were trained on a custom pedestrian dataset, with Tiny Yolov4 achieving 62% mean average precision (mAP). When deployed, this analysis indicates that some RCEPs are capable of achieving real-time pedestrian detection, benchmarked at 30FPS, with Tiny YOLOv4 achieving 40.48FPS on the NVIDIA Jetson Nano. The Raspberry Pi 4 paired with the Google Edge TPU also exceeded the real-time threshold, achieving 51.22FPS using MobileNetV2-SSD. Both models were quantized at FP16 and UINT8, which inherently have lower accuracy in favour of faster inferencing. Our analysis evaluates the real-time performance of the selected embedded devices when hosting lightweight CNNs, as well as offering numerous future research directions to improve the pedestrian detection.