Anchor Generation Optimization and Region of Interest Assignment for Vehicle Detection

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Faster R-CNN, a popular object detection method, faces challenges in specialized applications like onboard vehicle detection.
• Original RPNs have fixed receptive fields and uniform anchor generation, leading to suboptimal performance with varying object scales and perspective distortions.
• Existing Region of Interest (ROI) pooling methods lack precise control over feature map pixel counts, impacting detection accuracy.

Purpose of the Study:

• To enhance the accuracy and efficiency of Region Proposal Network (RPN)-based object detection, specifically for onboard vehicle detection.
• To address limitations in Faster R-CNN, including fixed receptive fields and inaccurate anchor generation due to perspective projection.
• To improve the control over feature map resolution in the ROI pooling layer for better detection performance.

Main Methods:

• Optimization of the original Faster R-CNN architecture, focusing on RPN and ROI pooling stages.
• Development of adaptive anchor generation strategies to account for object scale and position.
• Refinement of the ROI pooling mechanism for more accurate feature extraction from multiscale feature maps.

Main Results:

• Significant improvements in detection accuracy on the KITTI dataset for onboard vehicle detection.
• Demonstrated effectiveness without requiring complex parameter tuning or specialized training techniques.
• Validation of the method's applicability to other objects exhibiting foreshortening, such as pedestrians.

Conclusions:

• The proposed optimizations effectively address the limitations of standard Faster R-CNN for onboard object detection.
• The method offers a robust and adaptable solution applicable to various deep learning object detectors with multiscale features.
• This work provides a valuable advancement for real-world applications requiring accurate and efficient object detection in challenging scenarios.