A Comparative Study of Ego-centric and Cooperative Perception for Lane Change Prediction in Highway Driving Scenarios

Mozaffari, Sajjad; Arnold, Eduardo; Dianati, Mehrdad; Fallah, Saber · 2021 · Crossref

DOI: 10.5220/0010655700003061

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical challenge of lane change (LC) prediction for automated vehicles, specifically focusing on the limitations of egocentric perception sensors. While existing studies often assume full observability of traffic using top-down infrastructure cameras, real-world automated vehicles rely on onboard sensors (e.g., LiDAR, cameras) that suffer from limited range and occlusions. The authors investigate how these constraints degrade prediction performance and evaluate cooperative perception—where Connected Automated Vehicles (CAVs) share data via V2V communication—as a mitigation strategy. The study aims to quantify the performance gap between egocentric and cooperative perception and determine the benefits of varying CAV penetration rates. To conduct this comparative study, the authors developed two perception models to generate dataset variants from the highD naturalistic trajectory dataset, which originally provides full observability. The egocentric model simulates a 360-degree sensor with a defined range, marking pixels as unobservable if occluded by other vehicles or beyond the sensor range. The cooperative model extends this by assuming a percentage of vehicles are CAVs that share their observable fields, fusing these views via a logical OR operation. A Convolutional Neural Network (CNN) was trained to predict LC maneuvers (Left, Right, or Lane Keeping) using these generated datasets. The model utilized cropped images centered on the target vehicle, with observation and prediction windows of 1 second each. Experiments varied the time delay between observation and prediction (0 to 4 seconds) to assess short- and long-term prediction accuracy. The results demonstrate that egocentric perception significantly reduces observability to 63% compared to 100% in the full observability baseline. This limitation causes a notable drop in prediction accuracy, particularly as the prediction horizon increases. For instance, at a 4-second time delay, egocentric perception achieved 69.2% accuracy, compared to 73.49% for full observability. Cooperative perception with a 20% CAV penetration rate improved observability to 85% and narrowed the accuracy gap, achieving 72.28% at the 4-second delay. The study found that cooperative perception completely compensated for the performance gap at shorter delays (1–2 seconds) and reduced the gap to just 1% at the 4-second delay. Furthermore, increasing the CAV penetration rate linearly improved both observability and prediction accuracy, confirming that shared perception data effectively mitigates the occlusion and range limitations of individual sensors. The significance of this work lies in providing system designers with empirical evidence to weigh the costs and benefits of implementing cooperative perception systems. By quantifying the specific performance gains associated with V2V data sharing, the study highlights that while egocentric sensors are sufficient for short-term predictions, cooperative perception is crucial for reliable long-term maneuver prediction in congested highway scenarios. This contributes to the development of safer automated driving systems by addressing the realistic constraints of onboard sensing.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• situational awareness
• anticipation