Improved Mask R-CNN Multi-Target Detection and Segmentation for Autonomous Driving in Complex Scenes

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Deep Learning

Background:

• Current vision-based target detection and segmentation algorithms struggle with accuracy and mask quality in complex autonomous driving scenarios.
• Existing methods often exhibit limitations in effectively handling multi-target detection and segmentation.

Purpose of the Study:

• To improve the accuracy and segmentation quality of Mask R-CNN for autonomous driving applications.
• To address the challenges of multi-target detection and segmentation in complex traffic environments.

Main Methods:

• Enhanced Mask R-CNN by replacing ResNet with ResNeXt backbone for superior feature extraction.
• Integrated a bottom-up path enhancement strategy in Feature Pyramid Network (FPN) for improved feature fusion.
• Incorporated an efficient channel attention (ECA) module and replaced smooth L1 loss with CIoU loss for optimized performance.

Main Results:

• Achieved 62.62% mAP for target detection and 57.58% mAP for segmentation on the CityScapes dataset, outperforming original Mask R-CNN by 4.73% and 3.96% respectively.
• Demonstrated robust detection and segmentation across various traffic scenarios on the BDD autonomous driving dataset.

Conclusions:

• The proposed enhanced Mask R-CNN significantly improves performance in autonomous driving perception tasks.
• The model shows strong generalization capabilities across different datasets and complex traffic conditions.