A High-Performance Pedestrian Detector and Its Implementation on Embedded Systems for Hypermarket Environment

Enabling autonomous driving in hypermarket environments is a new challenge. The whole scenario is very different from traditional outdoor autonomous driving. To navigate in hypermarket environments, the vehicles need to know where all the surrounded pedestrians are. In addition, the detection model must be small enough to be executed in real-time speed on embedded systems. Therefore, we present a high-performance convolutional neural network for detecting moving indoor pedestrians as well as its implementation on embedded systems in this paper. The proposed CNN model can achieve the same high accuracy as YOLO v3 at the cost of only 27% of the original model size. When implemented on an embedded system, i.e., Jetson Xavier, this work achieves 30 fps @ 360p video format.

I. Introduction

Self-driving vehicles have gathered much attention recently. The massively produced vehicles available have achieved level 3+ autonomous driving capability already. However, to navigate in environment with crowds autonomously is still a challenging goal for self-driving vehicles. Such capability not only can be applied on indoor robots, but also is benefit to reaching the level 5 fully self-driving vehicles since they also need to face the traffic condition in crowded downtown. To navigate in hypermarket environments, the vehicles need to know where all the surrounded pedestrians are. Panoramic image processing has been matured to capture imaging information surrounding the vehicle. The key points of a successful indoor pedestrian detector therefore fall at the detection accuracy and the deployment on embedded systems.