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

Ci, Wenyan; Xuan, Jiayin; Lin, Runze; Lu, Shan · 2024 · DOAJ

DOI: 10.1109/ACCESS.2024.3359695

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 challenge of detecting road anomaly objects—uncommon or complex obstacles like dropped cargo or rocks—in Advanced Driving Assistance Systems (ADAS). Existing deep learning models are typically trained on known categories and struggle to generalize to unseen anomalies, often misclassifying them as background or failing to detect them entirely. This limitation poses significant safety risks in autonomous driving. The authors propose a novel method that fuses "assistant" techniques (uncertainty estimation and depth information) with "resynthesis" techniques (image reconstruction from semantic maps) to improve detection accuracy and reduce false positives caused by noise. The proposed framework operates through four main modules. First, a semantic segmentation network generates a semantic map and two uncertainty maps (softmax entropy and distance). Second, a synthesis module uses a conditional generative adversarial network (cGAN) to resynthesize an image from the semantic map. Anomalies are identified by calculating the perceptual difference between the original input and the synthesized image using a VGG-based feature extractor. Third, a depth module employs an RGB-D network with an Attention Feature Complementary mechanism to fuse RGB and depth features, enhancing geometric context. Fourth, a dissimilarity network integrates these features—original images, synthesized images, semantic maps, depth-enhanced maps, and uncertainty maps—to predict anomaly segmentation. Finally, a postprocessor applies superpixel segmentation and calculates an anomaly score for each region to refine localization and suppress false positives. Experiments were conducted on the RoadObstacle21 and Lost and Found datasets. The results demonstrate that the proposed multi-mechanism fusion method effectively detects both seen and unseen anomaly objects. By combining the complementary strengths of uncertainty/depth estimation and image resynthesis, the model mitigates the impact of unknown noise that typically plagues individual methods. The approach achieves higher detection rates and lower false positive rates compared to previous anomaly detection methods. The integration of depth information and uncertainty metrics into the dissimilarity network significantly enhances the model's ability to distinguish true anomalies from irrelevant scene elements. The significance of this work lies in its improved generalization ability for road anomaly detection. By moving beyond traditional classification-based approaches, the method provides a robust solution for identifying out-of-distribution objects in complex traffic environments. The proposed postprocessor further ensures precise localization, which is essential for safe autonomous navigation. This research contributes to the development of more reliable ADAS by addressing the critical gap in detecting unexpected road hazards that standard semantic segmentation models fail to recognize.

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.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• hazard perception