Vision-Based Collision Warning Systems with Deep Learning: A Systematic Review

Area of Science:

• Computer Vision
• Artificial Intelligence
• Automotive Safety

Background:

• Advanced driver assistance systems (ADAS) rely on timely collision prediction for warnings and emergency maneuvers.
• Vision-only deep learning collision warning systems offer a cost-effective alternative to multi-sensor approaches.
• Thorough assessment of vision-based systems is crucial for their reliable deployment.

Purpose of the Study:

• To systematically review and analyze ego-centric, vision-based collision warning systems utilizing deep learning.
• To identify deep learning techniques, evaluation datasets, experimental methodologies, and reported results.
• To assess the adequacy of experimental evaluations and identify challenges in the field.

Main Methods:

• Systematic literature search for vision-based, deep learning collision warning systems.
• Inclusion of 31 studies based on predefined criteria.
• Risk of bias assessment using PROBAST.
• Review of deep learning techniques, datasets, experiments, and results.

Main Results:

• Two main deep learning approaches were identified: direct collision probability prediction and pipeline-based threat metric computation.
• Most reviewed systems suffer from inadequate experimental evaluation, including a lack of quantitative experiments and biased datasets.
• The scarcity of suitable datasets presents a significant obstacle for robust system evaluation.

Conclusions:

• While deep learning shows potential for vision-based collision warning, current evaluation practices are insufficient.
• Addressing the lack of high-quality, unbiased datasets is critical for advancing the field.
• Future research should focus on developing standardized evaluation protocols and comprehensive datasets.