Lane Marking Detection via Deep Convolutional Neural Network.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Faster R-CNN demonstrates progress in object detection accuracy and efficiency.
• Existing Faster R-CNN models struggle with detecting small or tiny objects due to feature map size limitations.

Purpose of the Study:

• To develop a modified Faster R-CNN model for improved small object detection.
• To enhance the accuracy of detecting small objects in images and videos.

Main Methods:

• A fast, deep convolutional neural network based on a modified Faster R-CNN was developed.
• Key strategies include fast multi-level feature combination, context cues, and a novel anchor generation method.
• The algorithm was evaluated on the KITTI-ROAD dataset and a custom traffic scene lane markings dataset.

Main Results:

• The proposed algorithm significantly outperforms the standard Faster R-CNN in detecting small objects.
• Experiments confirmed improved accuracy for small object detection tasks.
• Performance was validated on both benchmark and custom datasets.

Conclusions:

• The modified Faster R-CNN effectively addresses the limitations of detecting small objects.
• The implemented strategies enhance the capability of deep learning models for precise small object recognition.
• This research contributes to more robust object detection systems, particularly in complex scenes.