Wide-View Sidewalk Dataset Based Pedestrian Safety Application

Chang Sun,Yong Guan,Wenpeng Ren,Jia Su

doi:10.1109/access.2019.2947165

Chang Sun, Yong Guan + Show 2 more

Open Access

https://doi.org/10.1109/access.2019.2947165

Copy DOI

Abstract

As the increasing number of people watching mobile phones while walking, sidewalk accidents occur more frequently. Using mobile camera instead of distracted pedestrians to monitor the road conditions ahead can effectively prevent pedestrian accidents. In order to develop this kind of applications, a wide-field sidewalk dataset becomes a necessity. In this paper, three major contributions are concluded. Firstly, a dataset quality evaluation model is proposed, which directs the establishment of a wide-view dataset named PESID for the sidewalk environment. PESID currently contains more than 1.9K labeled images which cover more than 5 districts, 10 communal facilities, 6 typical roads. Secondly, a criterion is presented to evaluate 9 up-to-date object detection algorithms in order to train a mobile feasible obstacle detection model. Finally, a reliable and low-cost framework is designed for obstacle detection based pedestrian safety application. The proposal is able to avoid 71.4% collisions on average by evaluating the mean average precision (MAP), the dangerous reminder omission rate and the false detection rate of the model.

Highlights

The number of pedestrian accidents is increasing rapidly recently
The lack of a high-quality sidewalk dataset is the main obstacle to the development of pedestrian safety applications
In order to establish a mobile device feasible obstacle detection based pedestrian safety application, methods are proposed to gather an enlarged camera field sidewalk dataset. We address this problem by building a widefield model for sidewalk dataset, through which it is found

Summary

INTRODUCTION

The number of pedestrian accidents is increasing rapidly recently. People who use smartphones while walking are called phubbers [1] and they are more stricken by a sidewalk accident [2]. In order to establish a mobile device feasible obstacle detection based pedestrian safety application, methods are proposed to gather an enlarged camera field sidewalk dataset. We address this problem by building a widefield model for sidewalk dataset, through which it is found. It is acceptable to use virtual dataset to evaluate the detection and segmentation algorithms, it cannot be used in practical security applications. It is found that the en-large camera field dataset based pedestrian safety application can help to avoid over 70% collisions with obstacles.

FRAMEWORK OVERVIEW

EVALUATION MODEL FOR SIDEWALK DATASET

MODEL TRAINING

EVALUATIONS

Findings

CONCLUSION AND FUTURE WORK