MDSF-YOLO: Advancing Object Detection With a Multiscale Dilated Sequence Fusion Network

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Systems

Background:

• Accurate traffic sign detection is crucial for autonomous driving safety.
• Existing methods struggle with varied sign scales, distances, and complex environments, leading to false positives and omissions.
• Increased sampling depth in current models exacerbates detection challenges.

Purpose of the Study:

• To develop a novel traffic sign detection framework, MDSF-YOLO, to overcome limitations of existing approaches.
• To enhance the precision of both localization and semantic information fusion in traffic sign recognition.
• To improve fine-grained feature extraction and optimize target localization and category identification.

Main Methods:

• Integration of multiscale sequence fusion (MSF) for synergistic feature integration across granularities.
• Introduction of a dilated-wise residual (DWR) module using dilated convolutions and channel-wise reparameterization.
• Implementation of a P2 detection head for shallow features and fully decoupled detection heads.

Main Results:

• MDSF-YOLO demonstrated superior performance over benchmark models on TT100K and CCTSDB2021 datasets, with mAP improvements of 8.8% and 2.4% respectively.
• The model significantly reduced false positives and leakage rates.
• Enhanced capabilities were verified on the VisDrone2019 dataset for drone-based object detection.

Conclusions:

• MDSF-YOLO offers an efficient and robust solution for traffic sign detection in autonomous driving.
• The proposed framework effectively addresses challenges posed by diverse sign scales and detection distances.
• The model provides a promising advancement for real-world autonomous driving applications and similar object detection tasks.