Probabilistic Prediction of Pedestrian Crossing Intention Using Roadside LiDAR Data

Junxuan Zhao,Yinfeng Li,Hongchao Liu,Hao Xu

doi:10.1109/access.2019.2927889

Abstract

Pedestrians are vulnerable road users that need proactive protection. While both autonomous and connected vehicle technologies aim to deliver greater safety benefits, current designs heavily rely on vehicle-based or on-board sensors and lack strategic real-time interactions with pedestrians who do not have any communication means. As pedestrians are passively protected by the system, they might be put into hazardous situations when vehicle-mounted sensors fail to detect their presence. This paper is part of ongoing research that uses roadside light detection and ranging (LiDAR) sensors to develop a human-in-the-loop system that brings pedestrians into the connected environment. To proactively protect pedestrians, accurate prediction of their intention for crossings at locations, such as unsignalized intersections and street mid-blocks is critical, and this paper presents a modified Naive Bayes approach for this purpose. It features a probabilistic approach to overcoming the common deficiencies in deterministic methods and provides valuable comparisons between feature-based data processing methods, such as artificial neural network (ANN) and model-based Naive Bayes approach. A case study was conducted by using a low-cost 16-line LiDAR sensor installed at the roadside. Pedestrians' crossing intention was predicted at a range of 0.5-3 s before actual crossings. The results satisfactorily demonstrated the properties of the modified Naive Bayes model, as well as its higher flexibility, compared with the ANN approaches in practice.

Highlights

Connected-vehicle technology will enable pedestrians, vehicles, roads, and infrastructures to communicate with each other and share vital traffic information through network technologies [1]
The results showed that Quantile Regression Forest (QRF) produced better results than Linear Quantile Regression (LQR) when the time-to-cross was less than 3 seconds
In the authors’ previous work [14], we trained a deep autoencoder-artificial neural network (DA-ANN) model using pedestrian trajectories extracted from roadside light detection and ranging (LiDAR) data to predict whether or not a pedestrian walking along the sidewalk will cross the road

Summary

INTRODUCTION

Connected-vehicle technology will enable pedestrians, vehicles, roads, and infrastructures to communicate with each other and share vital traffic information through network technologies [1]. In the authors’ previous work [14], we trained a deep autoencoder-artificial neural network (DA-ANN) model using pedestrian trajectories extracted from roadside LiDAR data to predict whether or not a pedestrian walking along the sidewalk will cross the road. The authors proposed a probabilistic model based on modified Naïve Bayes method and pedestrian trajectories extracted from roadside LiDAR sensors to predict pedestrian crossing intention before actual arrival at crossing facilities with real-time and quantitative confidence level information. PROPOSED PROBABILISTIC PREDICTION MODEL The proposed prediction model predicts crossing probabilities of pedestrians walking on a sidewalk or at an intersection corner in a time interval It was trained offline using pedestrians’ historical trajectory-level movement features, which were extracted from roadside LiDAR data. Results of two pedestrians and two vehicles in one sample data information of the same pedestrians and vehicles during a frame; Fig. 2(e) and Fig. 2(f) show the trajectories and speed three-second interval as an example

MODIFIED NAÏVE BAYES PREDICTION MODEL

19. Return

Findings

CONCLUSION