Lightweight Object Detection Ensemble Framework for Autonomous Vehicles in Challenging Weather Conditions.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Autonomous vehicles rely on computer vision for object detection, which is challenged by adverse weather conditions like fog and rain.
• Image corruption in poor weather significantly degrades object detection performance, impacting vehicle navigation and safety.

Purpose of the Study:

• To enhance the object detection capabilities of autonomous vehicles under adverse weather conditions.
• To improve the robustness and accuracy of self-driving car perception systems in challenging environments.

Main Methods:

• Ensembling multiple baseline deep learning models with various voting strategies for object detection.
• Utilizing data augmentation techniques to improve model performance, especially with limited training data.
• Leveraging transfer learning from baseline models to accelerate the object detection process.

Main Results:

• The ensembling approach demonstrated increased accuracy over baseline models in detecting objects in foggy and rainy conditions.
• Achieved 32.75% mean average precision (mAP) and 52.56% average precision (AP) for car detection in adverse weather.
• Validated the effectiveness of different voting strategies for bounding box predictions, enhancing explainability.

Conclusions:

• Ensembling deep learning models with data augmentation is an effective strategy for improving object detection in autonomous vehicles under adverse weather.
• The proposed methods offer a viable solution for resource-constrained autonomous systems operating in uncertain weather conditions.
• The techniques enhance the reliability and safety of self-driving cars by improving perception accuracy in challenging environments.