Salient Object Detection in Traffic Scene Through the TSOD10K Dataset

Area of Science:

• Computer Vision
• Artificial Intelligence
• Intelligent Transportation Systems

Background:

• Traditional Salient Object Detection (SOD) focuses on visual distinctiveness, which is insufficient for traffic safety.
• Traffic Salient Object Detection (TSOD) requires identifying semantically critical objects for driver attention, even with low visual contrast.
• Existing benchmarks lack the specificity needed for TSOD in diverse and challenging traffic conditions.

Purpose of the Study:

• To establish a large-scale dataset (TSOD10K) for Traffic Salient Object Detection.
• To develop a novel Mamba-based model (Tramba) for accurate TSOD.
• To create a benchmark for evaluating TSOD models in intelligent transportation systems.

Main Methods:

• Collected and annotated the first large-scale TSOD dataset, TSOD10K, featuring diverse traffic scenarios and conditions.
• Proposed Tramba, a Mamba-based TSOD model incorporating a Dual-Frequency Visual State Space module for enhanced detail perception.
• Introduced a Helix 2D-Selective-Scan (Helix-SS2D) mechanism to integrate driving attention priors and capture spatial dependencies.

Main Results:

• Tramba demonstrated superior performance compared to 25 existing Natural Scene Image SOD models on the TSOD10K dataset.
• The proposed Dual-Frequency Visual State Space module effectively enhances perception of fine details and global structures.
• The Helix-SS2D mechanism successfully emphasized critical regions by incorporating driving attention priors.

Conclusions:

• The research establishes a foundational benchmark and model for safety-aware saliency analysis in intelligent transportation.
• TSOD10K and Tramba provide crucial resources for advancing autonomous and assisted driving systems.
• This work re-defines saliency for driving contexts by bridging perception and contextual risk assessment.