Impact of Perception Errors in Vision-Based Detection and Tracking Pipelines on Pedestrian Trajectory Prediction in...

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Pedestrian trajectory prediction is vital for autonomous driving collision avoidance.
• Traditional methods rely on accurate pedestrian detection and tracking, but real-world inaccuracies are often overlooked in research.
• Existing trajectory prediction models are typically evaluated on public datasets, neglecting cumulative errors from upstream processes.

Purpose of the Study:

• To investigate the cumulative impact of errors from integrated detection, tracking, and trajectory prediction pipelines.
• To empirically analyze errors introduced at each stage of the pipeline and their collective effect on prediction accuracy.
• To evaluate these integrated models using custom datasets collected in Taiwan.

Main Methods:

• Developed an integrated pipeline combining pedestrian detection, tracking, and trajectory prediction.
• Conducted empirical analysis to quantify errors at each stage (detection, tracking, prediction).
• Evaluated model performance on custom datasets capturing real-world conditions.

Main Results:

• Detection and tracking errors significantly influence downstream trajectory prediction accuracy.
• The cumulative effect of errors across the pipeline was empirically demonstrated.
• Performance was assessed across diverse custom datasets, highlighting practical implications.

Conclusions:

• Real-world detection and tracking inaccuracies are critical factors affecting autonomous driving safety systems.
• Future research must consider the end-to-end pipeline performance, not just isolated prediction models.
• Addressing upstream error propagation is essential for robust pedestrian trajectory prediction.