Object Detection Based on Lightweight YOLOX for Autonomous Driving

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Autonomous driving systems require rapid and accurate target detection for safe operation.
• Existing models often struggle with complex driving scenarios, leading to safety risks.
• Lightweight and efficient models are crucial for real-time application in driver assistance systems.

Purpose of the Study:

• To develop an improved object detection model for autonomous driving.
• To enhance the accuracy and speed of target detection in complex scenes.
• To create a lightweight model suitable for real-time applications.

Main Methods:

• Proposed a novel lightweight feature extraction model, ShuffDet, to replace CSPDark53 in the YOLOX algorithm.
• Integrated an attention mechanism into the Path Aggregation Feature Pyramid Network (PAFPN) to improve information focus.
• Combined these enhancements into the ShuffYOLOX model for autonomous driving applications.

Main Results:

• ShuffYOLOX achieved a mean average precision (mAP) of 92.20% on the KITTI dataset.
• Reduced model parameters by 34.57% and GFLOPS by 42.19% compared to the original network.
• Increased Frames Per Second (FPS) by 65%, indicating significant speed improvements.

Conclusions:

• ShuffYOLOX offers a superior balance of accuracy and efficiency for autonomous driving.
• The model's lightweight nature and enhanced performance make it highly suitable for real-time driver assistance systems.
• This research contributes to the advancement of safer and more responsive autonomous vehicles.