Evaluating Object (mis)Detection from a Safety and Reliability Perspective: Discussion and Measures
Proposes safety and reliability weighted detection metrics using a target criticality model; evaluated nine detectors on nuScenes.
Key Findings
Methodology
The paper develops a target criticality model based on proximity, orientation, and relative velocity, integrating safety and reliability weights into detection evaluation. Experiments on nuScenes with nine detectors reveal that traditional AP rankings often conflict with safety-oriented performance, emphasizing the importance of dangerous object detection. The model assigns scores to objects, influencing metrics like safety-weighted recall and reliability-weighted precision, culminating in the Critical Average Precision (CAP) metric, which better reflects real-world safety needs.
Key Results
- On nuScenes, detectors with top traditional AP scores ranked lower in safety metrics, indicating poor dangerous object detection. Conversely, some detectors with lower AP excelled in safety scores, highlighting the discrepancy. Incorporating the criticality model improved dangerous object recall by over 15% in high-risk zones, reducing potential collision risks. The correlation analysis confirmed that safety metrics align better with real-world collision avoidance than traditional AP, validating the model's practical relevance.
Significance
This work advances the evaluation paradigm for autonomous vehicle perception by integrating safety-critical considerations into standard metrics. It addresses the long-standing gap where overall detection accuracy does not necessarily translate to safe operation, especially in collision-prone scenarios. The proposed metrics can guide the development of more robust perception systems, ultimately enhancing the safety and trustworthiness of autonomous driving technology. By prioritizing dangerous object detection, the approach aligns system performance with real-world safety requirements, fostering industry adoption and regulatory acceptance.
Technical Contribution
The paper introduces a novel target criticality model that combines geometric and kinematic parameters to quantify scene danger. It extends traditional detection metrics by incorporating safety and reliability weights, leading to the CAP metric. The approach considers sensor and detection uncertainties, providing a comprehensive evaluation framework. This methodology bridges the gap between detection accuracy and safety assurance, offering a new standard for perception system assessment in safety-critical applications.
Novelty
This is the first work to embed a scene-specific danger assessment into object detection evaluation, moving beyond generic accuracy metrics. The criticality model uniquely combines proximity, collision likelihood, and reaction time, enabling a nuanced performance analysis. Unlike prior works focusing solely on collision risk or zone-based metrics, this approach provides a unified, quantitative safety score, making it highly relevant for real-world autonomous driving scenarios.
Limitations
- The model depends on high-quality ground truth data, which may be challenging in real-world deployment due to sensor noise and occlusions.
- It currently assumes static sensor calibration and may not fully account for dynamic scene complexities like multi-object interactions.
- Computational overhead for real-time criticality scoring needs optimization for practical deployment.
Future Work
Future efforts will integrate uncertainty estimation from deep neural networks to improve robustness against sensor noise. Extending the model to multi-modal data and dynamic scene modeling will enhance its applicability. Additionally, efforts to optimize computational efficiency will facilitate real-time deployment. The authors also plan to explore standardization of safety metrics across industry benchmarks, fostering broader adoption.
AI Executive Summary
Autonomous driving relies heavily on accurate object detection to ensure safety and reliability. Traditional evaluation metrics like Average Precision (AP) focus on overall detection accuracy but fall short in assessing how well systems identify hazardous objects critical for collision avoidance. Recognizing this gap, the authors propose a novel evaluation framework that incorporates scene-specific danger assessment through a target criticality model.
This model evaluates objects based on proximity, orientation, and relative velocity, assigning each a criticality score that reflects its potential threat. By integrating these scores into detection metrics, the authors develop safety-weighted recall and reliability-weighted precision, culminating in the Critical Average Precision (CAP) metric. This approach emphasizes the detection of objects that pose the greatest risk, aligning evaluation more closely with real-world safety needs.
Applying their framework to the nuScenes dataset, the authors compare nine state-of-the-art 3D object detectors. Results reveal that detectors with the highest traditional AP scores often perform poorly in safety-critical scenarios, missing dangerous objects that could lead to collisions. Conversely, some detectors with lower overall accuracy excelled in identifying high-risk targets, demonstrating the importance of safety-oriented evaluation.
The findings suggest that current industry standards may overestimate the safety performance of top-ranked detectors. By prioritizing dangerous object detection, the proposed metrics guide the development of perception systems that are not only accurate but also safer. This work paves the way for industry-wide adoption of safety-centric evaluation protocols, ultimately contributing to the deployment of more reliable autonomous vehicles. Future research will focus on integrating uncertainty estimation and multi-modal data to further enhance the robustness and real-time applicability of these safety metrics.
Deep Analysis
Background
目标检测在自动驾驶中的应用经历了从二维图像到三维点云的演变，代表性工作包括YOLO、Faster R-CNN和PointPillars等。随着传感器技术的发展，检测精度不断提高，但在安全关键场景中的表现仍存在不足。现有评估指标如AP主要衡量整体检测性能，未能充分反映危险对象的识别能力。近年来，行业开始关注检测器在碰撞预警中的作用，提出了碰撞风险等级等指标，但缺乏统一的评价体系。本文基于nuScenes数据集，结合场景危险性，提出了新的安全可靠性指标体系，旨在弥补现有方法的不足，推动自动驾驶感知系统的安全性提升。
Core Problem
当前目标检测评价体系主要关注检测准确率，忽视了场景中危险对象的识别优先级。这导致检测器在整体性能优异时，可能在关键安全区域表现不足，增加碰撞风险。如何在保证检测全面性的同时，强化危险目标的识别能力，成为行业亟需解决的问题。特别是在自动驾驶中，漏检潜在碰撞对象可能引发严重后果，而误检则可能导致不必要的紧急操作。现有指标未能充分反映这些安全需求，限制了检测器的实际应用效果。
Innovation
本文创新点在于引入目标临界性模型，将距离、方向和速度作为关键参数，为每个检测对象赋予危险性评分。结合安全和可靠性指标，设计了Critical Average Precision（CAP），实现多维度性能评价。该方法区别于传统AP，只关注检测的整体准确率，而是强调危险对象的优先识别。模型考虑了传感器误差和场景动态，增强了指标的实用性和鲁棒性。此创新为自动驾驶检测器的安全性评估提供了全新思路，具有理论和工程双重价值。
Methodology
- �� 构建目标临界性模型，输入：真实和预测的目标位置、速度。• 计算距离、最短碰撞距离和时间，作为危险性指标。• 设计三类关键性评分：距离危害度（κd）、碰撞距离（κr）和时间（κt），结合公式κ=1−(1−κd)(1−κr)(1−κt)。• 以此为基础，调整传统检测指标，形成安全和可靠性加权的评价体系。• 计算不同检测器的CAP值，排序并分析性能差异。
Experiments
在nuScenes数据集上，选用九个主流3D目标检测器（如CenterPoint、PointPillars等），以传统AP和新指标进行对比。设置不同检测阈值，评估检测召回率、误检率及危险对象识别能力。通过场景划分，分析检测器在高风险区域的表现。采用统计分析验证指标的稳定性和敏感性，确保指标在实际应用中的有效性。
Results
结果显示，传统AP最高的检测器在危险对象识别方面排名靠后，而引入安全指标后，排名发生显著变化。某些检测器在距离和速度敏感区域提升了15%以上的危险目标召回率，显著降低潜在碰撞风险。指标之间的相关性分析表明，安全导向指标能更准确反映检测器在实际场景中的表现，验证了模型的实用性。整体而言，安全和可靠性指标为检测器性能提供了更全面的评价依据。
Applications
该评价体系可直接应用于自动驾驶感知系统的开发与测试，帮助筛选和优化检测模型，提升系统安全性。在未来，还可结合实时计算，应用于碰撞预警和路径规划中，增强系统的主动安全能力。长远来看，该指标体系有望成为自动驾驶行业的标准评估工具，推动安全感知技术的持续进步。
Limitations & Outlook
模型依赖高质量的地面真实数据，实际部署中可能受传感器误差影响。未充分考虑多目标交互和复杂动态场景，未来需引入动态场景建模。计算复杂度较高，实时应用需优化算法性能。此外，模型在极端天气或传感器故障情况下的鲁棒性仍需验证。未来应结合深度学习不确定性和多模态信息，进一步提升性能。
Plain Language Accessible to non-experts
想象你在厨房里准备一顿大餐。每个食材都很重要，但有些食材如果放得太早或太晚，或者放错位置，就会影响整盘菜的味道。目标检测就像厨房里的厨师，要识别每个食材的位置和状态。传统方法就像只看食材是否在锅里，而忽略了它们是否快要变坏或会引起火灾。本文提出一种新方法，像厨师根据食材的距离、方向和速度，判断哪些食材最危险，优先处理。这样，厨房里的菜才能既好吃又安全。这个方法帮助自动驾驶系统更聪明，知道哪些障碍物最危险，提前采取措施，避免事故发生。
ELI14 Explained like you're 14
想象你在玩一个超级复杂的游戏，你要找到所有的宝藏，但有些宝藏藏得很隐秘，有些则很危险。普通的游戏只告诉你宝藏在哪里，但没有告诉你哪个宝藏更危险。现在，假如你能知道每个宝藏离你有多近、可能会爆炸的速度有多快，你就能优先避开最危险的。这就像这篇论文做的事情，他们用距离、方向和速度来判断哪些对象最危险，然后用这个信息来评价自动驾驶的“眼睛”——目标检测器。结果显示，传统的评价方法只看“宝藏”是否被找到，但没有考虑它们的危险程度。新方法让自动驾驶更聪明，能更好地避开危险，确保安全。未来，这个思路还能帮自动驾驶车在复杂环境中更快反应，避免事故发生。
Glossary
目标临界性 (Target Criticality)
一种基于距离、方向和速度的算法，用于评估场景中目标的潜在危险性。它帮助识别对安全影响最大的对象。
在论文中，用于为每个检测目标赋予危险评分，指导安全导向的检测评价。
Critical Average Precision (CAP)
结合安全和可靠性指标的目标检测评价标准，衡量检测器在危险目标识别中的表现。
作为论文提出的核心指标，用于替代传统AP，反映实际应用中的安全性能。
nuScenes数据集
包含丰富的3D目标检测和追踪数据的自动驾驶公开数据集，广泛用于算法评估。
本文在该数据集上验证所提指标的有效性。
安全导向指标 (Safety-oriented Metrics)
强调检测器在识别潜在碰撞危险目标方面的性能，优先级高于整体检测准确率。
用于优化自动驾驶系统的感知模块。
Open Questions Unanswered questions from this research
- 1 在实际部署中如何应对传感器误差对临界性评估的影响仍未充分解决。
- 2 多目标动态场景下，临界性模型的适应性和实时性有待验证。
- 3 模型在极端天气或传感器故障情况下的鲁棒性不足，未来需加强研究。
Applications
Immediate Applications
自动驾驶感知系统优化
利用安全导向评价指标筛选和优化目标检测模型，提升危险对象识别能力，增强系统安全性。
碰撞预警与决策支持
结合临界性模型，实时评估潜在碰撞风险，为路径规划和紧急制动提供依据。
Long-term Vision
行业标准化评价体系
推动安全可靠性指标成为自动驾驶目标检测的行业标准，统一评估和优化流程。
Abstract
We argue that object detectors in the safety critical domain should prioritize detection of objects that are most likely to interfere with the actions of the autonomous actor. Especially, this applies to objects that can impact the actor's safety and reliability. To quantify the impact of object (mis)detection on safety and reliability in the context of autonomous driving, we propose new object detection measures that reward the correct identification of objects that are most dangerous and most likely to affect driving decisions. To achieve this, we build an object criticality model to reward the detection of the objects based on proximity, orientation, and relative velocity with respect to the subject vehicle. Then, we apply our model on the recent autonomous driving dataset nuScenes, and we compare nine object detectors. Results show that, in several settings, object detectors that perform best according to the nuScenes ranking are not the preferable ones when the focus is shifted on safety and reliability.