Robust Traffic Light Detection Using Salience-Sensitive Loss: Computational Framework and Evaluations
DOI: 10.1109/iv55152.2023.10186624
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 safety challenge of accurately detecting traffic lights relevant to an autonomous vehicle’s immediate driving decisions. In complex driving scenes, multiple traffic lights may be present, but only those influencing the ego-vehicle’s next maneuver are pertinent. The authors define these as "salient" lights, distinguishing them from non-salient lights that do not affect the current trajectory. To tackle this, the study introduces the LAVA Salient Lights Dataset, the first US-based traffic light dataset annotated with a salience property, and proposes a computational framework using a Salience-Sensitive Focal Loss to improve detection performance on these critical objects. The methodology involves collecting 30,566 traffic light annotations from the greater San Diego area, comprising 9,051 salient and 21,515 non-salient instances. Salience is defined based on the ego-vehicle’s lane and intended maneuver; for example, a straight-ahead light is salient for a vehicle driving straight, while left-turn lights are non-salient in that context. The authors employ the Deformable DETR transformer model for object detection, utilizing a ResNet-50 backbone. They compare two training approaches: a standard Deformable DETR and one modified with a custom Salience-Sensitive Focal Loss. This loss function introduces a weighting factor ($\omega_{SL}$) that increases the penalty for misclassifying salient lights, thereby forcing the model to prioritize accuracy on these critical detections. The dataset was split into 80% training, 10% validation, and 10% testing sets, with models trained for 50 epochs. Experimental results demonstrate that the Salience-Sensitive Focal Loss significantly enhances detection performance. Precision-recall curves indicate that the model trained with the salience-sensitive loss achieves consistently higher recall for salient traffic lights compared to the standard model, even at high precision thresholds. Furthermore, the modified model maintains stronger recall for all traffic lights overall. Analysis of confidence scores reveals that at high confidence levels, the salience-sensitive model retains robust detection of salient lights, whereas the standard model’s performance drops more sharply. This confirms that the loss function successfully prioritizes the detection of lights critical to the driver’s decision-making process. The significance of this work lies in its contribution to safer autonomous driving systems by ensuring that detectors focus on the most relevant environmental cues. By introducing the first US dataset with salience annotations and validating the effectiveness of salience-aware loss functions, the paper provides a framework for improving the reliability of traffic light detection. The findings suggest that incorporating semantic relevance into the training process allows transformer-based detectors to better handle complex scenes with multiple, potentially distracting traffic signals, ultimately supporting more accurate trajectory planning and decision-making for autonomous vehicles.
Key finding
Training a Deformable DETR model with a Salience-Sensitive Focal Loss function significantly improves the recall of salient traffic lights compared to training with standard focal loss.
Methodology
lab_experiment
Sample size: 30566
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. Discovered via author_sweep_intake on 2026-05-28.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | author_sweep | — | — | 2 | 2026-05-28 |
| archive | success | canonical_url | — | — | 1 | 2026-06-06 |
| extract | success | cached | — | — | 3 | 2026-06-10 |
| clean | success | clean | — | — | 1 | 2026-06-04 |
| chunk | success | chunk | — | — | 1 | 2026-06-04 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-04 |
| enrich | success | — | — | — | 1 | 2026-05-28 |
| promote | success | — | — | — | 1 | 2026-06-04 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 2 | 2026-06-10 |
| tag | success | vector_similarity | — | — | 15 | 2026-06-11 |
| verify | success | — | — | — | 2 | 2026-06-10 |
Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.
Topics
Ranked by relevance to this paper. Hover a topic for its definition.