Abstract:The detection of unknown traffic obstacles is vital to ensure safe autonomous driving. The standard object-detection methods cannot identify unknown objects that are not included under predefined categories. This is because object-detection methods are trained to assign a background label to pixels corresponding to the presence of unknown objects. To address this problem, the pixel-wise anomaly-detection approach has attracted increased research attention. Anomaly-detection techniques, such as uncertainty estimation and perceptual difference from reconstructed images, make it possible to identify pixels of unknown objects as out-of-distribution (OoD) samples. However, when applied to images with many unknowns and complex components, such as driving scenes, these methods often exhibit unstable performance. The purpose of this study is to achieve stable performance for detecting unknown objects by incorporating the object-detection fashions into the pixel-wise anomaly detection methods. To achieve this goal, we adopt a semantic-segmentation network with a sigmoid head that simultaneously provides pixel-wise anomaly scores and objectness scores. Our experimental results show that the objectness scores play an important role in improving the detection performance. Based on these results, we propose a novel anomaly score by integrating these two scores, which we term as unknown objectness score. Quantitative evaluations show that the proposed method outperforms state-of-the-art methods when applied to the publicly available datasets.

What problem does this paper attempt to address?

This paper attempts to address the problem of detecting unknown traffic obstacles in autonomous driving scenarios. Specifically, standard object detection methods cannot recognize unknown objects that are not included in predefined categories, as these methods typically label the pixels of unknown objects as background. This leads to unstable performance when detecting unknown objects in complex driving scenarios. To solve this problem, the authors propose a method that combines pixel-level anomaly detection and object detection by introducing a new anomaly score—unknown objectness score, to improve the stability of detecting unknown objects. ### Summary of the Core Content of the Paper: 1. **Problem Background**: - In autonomous driving, detecting various obstacles on the road is crucial for ensuring driving safety. - Standard object detection methods cannot recognize unknown objects of undefined categories because these methods label the pixels of unknown objects as background. - Although pixel-level anomaly detection methods can identify unknown regions in images, their performance is unstable in complex driving scenarios and prone to false positives. 2. **Solution**: - The authors propose a method based on a semantic segmentation network that uses a sigmoid head, capable of providing both pixel-level anomaly scores and objectness scores. - By combining these two scores, a new score—unknown objectness score is defined to improve the stability of detecting unknown objects. 3. **Method Details**: - **Unknown Scores**: Defined by calculating the product of the prediction probabilities of all predefined categories. - **Objectness Scores**: Defined by merging predefined foreground categories into one object category and calculating its prediction probability as the objectness score. - **Unknown Objectness Scores**: Defined by combining unknown scores and objectness scores to reduce false positives in background regions. 4. **Experimental Results**: - The authors conducted experiments on multiple public datasets, and the results show that the proposed method outperforms existing state-of-the-art methods in detecting unknown road obstacles. - It performs particularly well in reducing false positives in background regions. 5. **Contributions**: - Proposed a new road obstacle detection method based on a new anomaly score—unknown objectness score. - Experimental results show that this method outperforms existing methods on public datasets. - The method is based solely on a semantic segmentation network, requiring no additional modules, facilitating rapid detection. ### Conclusion: This study proposes a new method that combines pixel-level anomaly detection and object detection by introducing the unknown objectness score, improving the stability and accuracy of detecting unknown objects in complex driving scenarios. Experimental results validate the effectiveness of this method, which is expected to enhance the safety of autonomous driving systems.

Road Obstacle Detection based on Unknown Objectness Scores

Probabilistic Approach for Road-Users Detection

A Novel Method for Road Anomaly Objects Detection in the Traffic Environment With Multi-Mechanism Fusion

Anomaly Candidate Identification and Starting Time Estimation of Vehicles from Traffic Videos.

UGainS: Uncertainty Guided Anomaly Instance Segmentation

Uncertainty Evaluation of Object Detection Algorithms for Autonomous Vehicles

Uncertainty Aware Proposal Segmentation for Unknown Object Detection

Roadside Unit-based Unknown Object Detection in Adverse Weather Conditions for Smart Internet of Vehicles

UNCOVER: Unknown Class Object Detection for Autonomous Vehicles in Real-time

Anomaly Detection in Traffic Scenes via Spatial-aware Motion Reconstruction

Uncertainty-Aware Point-Cloud Semantic Segmentation for Unstructured Roads

Unknown Sniffer for Object Detection: Don't Turn a Blind Eye to Unknown Objects

An Objectness Score for Accurate and Fast Detection during Navigation

Object and Patch Based Anomaly Detection and Localization in Crowded Scenes

Road Obstacle Object Detection Based on Improved YOLO V4

Object-Oriented Anomaly Detection In Surveillance Videos

Prediction Surface Uncertainty Quantification in Object Detection Models for Autonomous Driving

A novel real-time object detection method for complex road scenes based on YOLOv7-tiny

Anomaly-Aware Semantic Segmentation by Leveraging Synthetic-Unknown Data

Unsupervised Road Anomaly Detection with Language Anchors.

SegmentMeIfYouCan: A Benchmark for Anomaly Segmentation