Saleh, Khaled (The University of Newcastle), Grigorev, Artur (University of technology Sydney), Mihaita, Adriana-Simona (University of Technology in Sydney)

Cyclist Near-Miss Detection Using Lightweight Deep Temporal Neural Networks

Scheduled for presentation during the Regular Session "S32b-AI-Driven Traffic Monitoring, Safety, and Anomaly Detection" (FR-EA-T32), Friday, November 21, 2025, 14:50−15:30, Southport 2

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Near-miss incidents, where cyclists narrowly avoid collisions, are critical for understanding and improving urban cycling safety but are often underreported in official statistics. This paper introduces XceptionCycle, a lightweight and efficient deep neural network designed to detect cyclist near-miss incidents using time-series data from Inertial Measurement Units (IMUs) and GPS sensors. Building on the XceptionTime architecture, our model incorporates inverted bottlenecks and multi-scale depthwise separable convolutions to extract rich temporal features while maintaining a low computational footprint. We benchmark XceptionCycle on the large-scale SimRa dataset, demonstrating superior discriminative performance with an 81.99% Area Under the ROC Curve and strong robustness with a 48.33% Matthews correlation coefficient, outperforming state-of-the-art models, while requiring less than half the number of trainable parameters. These results highlight XceptionCycle's potential for real-time near-miss detection and its suitability for resource-constrained environments such as mobile safety applications.