Real and Pseudo Pedestrian Detection Method with CA-YOLOv5s Based on Stereo Image Fusion

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Deep learning models, particularly convolutional neural networks (CNNs), have advanced pedestrian detection.
• Pseudo pedestrians (e.g., reflections, shadows) degrade the accuracy of existing detection systems.
• Distinguishing real from pseudo pedestrians remains a challenge in computer vision.

Purpose of the Study:

• To develop a robust method for differentiating real pedestrians from pseudo ones.
• To improve the accuracy and reliability of pedestrian detection systems.
• To address the limitations of current algorithms in handling false positives.

Main Methods:

• Utilized binocular stereo cameras to capture two-view pedestrian images.
• Applied a proposed CA-YOLOv5s algorithm for pedestrian detection in both left and right views.
• Integrated stereo image fusion with 3D spatial coordinate calculation (Zhengyou Zhang's method) and RANSAC plane-fitting for feature extraction.
• Employed a Support Vector Machine (SVM) classifier trained on extracted 3D features for final real/pseudo pedestrian classification.

Main Results:

• The proposed CA-YOLOv5s based stereo image fusion method effectively identifies and filters out pseudo pedestrians.
• Achieved significant improvements in pedestrian detection accuracy and precision compared to existing algorithms.
• Demonstrated superior performance on datasets containing both real and pseudo pedestrian instances.

Conclusions:

• The developed method offers a viable solution for the pseudo pedestrian detection problem in computer vision.
• Stereo image fusion combined with deep learning and 3D analysis enhances the robustness of pedestrian detection.
• This approach holds promise for applications requiring high-accuracy pedestrian identification, such as autonomous driving and surveillance.