STRIDE: Street View-based Environmental Feature Detection and Pedestrian Collision Prediction

TL;DR

A baseline method is proposed that incorporates a collision prediction module into a state-of-the-art detection model to tackle both tasks simultaneously and demonstrates a significant correlation between object detection of built environment elements and pedestrian collision frequency prediction.

Abstract

This paper introduces a novel benchmark to study the impact and relationship of built environment elements on pedestrian collision prediction, intending to enhance environmental awareness in autonomous driving systems to prevent pedestrian injuries actively. We introduce a built environment detection task in large-scale panoramic images and a detection-based pedestrian collision frequency prediction task. We propose a baseline method that incorporates a collision prediction module into a state-of-the-art detection model to tackle both tasks simultaneously. Our experiments demonstrate a significant correlation between object detection of built environment elements and pedestrian collision frequency prediction. Our results are a stepping stone towards understanding the interdependencies between built environment conditions and pedestrian safety.

Figures (6)

• Figure 1: STRIDE. Given specific city coordinates (left) and the corresponding panoramic street view image (middle), we propose to predict the number of pedestrian collisions in those coordinates (right) by detecting built environment features (middle).
• Figure 2: Number of labeled bounding boxes (y-axis) per class in each fold and their corresponding category name (x-axis).
• Figure 3: Distribution of the number of bounding box annotations per image. The figure shows the number of images (y-axis) with a certain number of annotated instances (x-axis). Our images contain a varying range of instances with a similar distribution among folds.
• Figure 4: Number of annotated bounding boxes (y-axis) per relative box area interval (x-axis). Most instances in our dataset have small relative sizes due to the large scale of our images.
• Figure 5: Pedestrian Collision frequency distribution among training folds (a) and testing set (b). Figures portray the percentage of images (y-axis) for each amount of pedestrian collisions (x-axis). Our dataset maintains a constant long-tail distribution among the training folds and the testing set. Note: The figures were cut to a maximum of 30 collisions for better visualization.