Faster R-CNN for Robust Pedestrian Detection Using Semantic Segmentation Network.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) excel at feature representation for tasks like pedestrian detection.
• A key challenge in CNN-based pedestrian detection is reducing false positives caused by hard negative samples (e.g., foliage, traffic signals).
• High-level semantic vision cues can help distinguish pedestrians from similar-looking background objects.

Purpose of the Study:

• To propose a region-based CNN method that leverages semantic cues to enhance pedestrian detection accuracy.
• To address the issue of false positives on hard negative samples in pedestrian detection systems.
• To improve the robustness and reliability of pedestrian detection, especially at varying scales.

Main Methods:

• An extension of the Faster R-CNN framework incorporating a semantic image segmentation network branch.
• Integration of complementary semantic features with standard convolutional features.
• Utilization of multi-resolution feature maps from various network layers for scale invariance.
• Employing a boosted forest classifier for cascaded training and hard negative mining.

Main Results:

• The proposed semantic network integration demonstrates improved pedestrian detection accuracy on the Caltech pedestrian dataset.
• The method effectively utilizes semantic cues to mitigate false positives from challenging background elements.
• Robust detection performance was achieved using the deep VGG16 model within the proposed framework.

Conclusions:

• Integrating semantic image segmentation into region-based CNNs significantly enhances pedestrian detection performance.
• The proposed method offers a robust solution for reducing false positives in complex scenes.
• This approach advances the state-of-the-art in pedestrian detection by effectively handling difficult negative samples.