PVformer: Pedestrian and Vehicle Detection Algorithm Based on Swin Transformer in Rainy Scenes

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Deep Learning

Background:

• Accurate pedestrian and vehicle detection is crucial for autonomous vehicle safety.
• Transformer-based object detection shows promise but struggles with adverse weather like rain.
• Rainfall significantly degrades image quality, challenging current detection algorithms.

Purpose of the Study:

• To develop an end-to-end algorithm for robust pedestrian and vehicle detection in rainy conditions.
• To improve the accuracy and reliability of object detection systems for autonomous driving in adverse weather.
• To mitigate the impact of rain streak occlusion on detection performance.

Main Methods:

• Proposed PVformer, an algorithm based on the Swin Transformer with an integrated deraining module.
• Introduced a four-branch feature mapping model using Transformer blocks for single-image deraining.
• Designed a local enhancement perception block combining CNN and Transformer to address small object detection challenges.
• Employed transfer learning to train the combined deraining and detection modules.

Main Results:

• The PVformer algorithm demonstrated strong performance in rainy scenarios.
• Significantly improved the accuracy of pedestrian and vehicle detection under rainy conditions.
• The integrated deraining module effectively reduced the influence of rain streak occlusion.

Conclusions:

• The proposed PVformer algorithm enhances image quality and detection accuracy in rainy scenes for autonomous driving.
• The novel approach effectively addresses the challenges of object detection in adverse weather.
• PVformer offers a promising solution for improving the safety and reliability of self-driving vehicles.