Pedestrian detection using a translation-invariant wavelet residual dense super-resolution.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Pedestrian detection is crucial for applications like autonomous driving and surveillance.
• Existing methods struggle with low-resolution images, hindering real-world performance.
• Low-resolution imaging devices remain prevalent, necessitating robust detection solutions.

Purpose of the Study:

• To develop an effective method for enhancing low-resolution images for improved pedestrian detection.
• To address the limitations of current pedestrian detection techniques in low-resolution scenarios.
• To propose a novel super-resolution technique integrated with object detection.

Main Methods:

• A novel end-to-end Translation-invariant Wavelet Residual Dense Super-Resolution (TiWRD-SR) method is proposed.
• Stationary Wavelet Transform (SWT) decomposes images into low- and high-frequency components for targeted reconstruction.
• A high-to-low branch information transmission (H2LBIT) mechanism and a novel loss function are introduced to enhance structural details.

Main Results:

• The TiWRD-SR method successfully upscales low-resolution (LR) images to super-resolution (SR) images.
• The enhanced SR images improve the performance of pedestrian detection using Yolov4.
• Experimental results demonstrate significant improvements in detection accuracy on low-resolution images.

Conclusions:

• The proposed TiWRD-SR method effectively enhances image quality for better pedestrian detection.
• Integrating super-resolution with object detection offers a viable solution for low-resolution challenges.
• The method shows promise for real-world applications requiring reliable pedestrian detection in degraded image conditions.