A comprehensive swarming intelligent method for optimizing deep learning-based object detection by unmanned ground...

Area of Science:

• Artificial Intelligence
• Computer Vision
• Robotics

Background:

• Unmanned ground vehicles (UGVs) rely on artificial intelligence for environmental perception.
• Deep learning object detection, particularly Faster R-CNN, is crucial for UGV perception.
• The region proposal network (RPN) in Faster R-CNN has limitations in its exploration space.

Purpose of the Study:

• To develop an improved region proposal network (RPN) called Boosted RPN (BRPN) for enhanced object detection in UGVs.
• To address the expression limitations of the standard RPN.
• To improve the adaptability and classification capacity of object detection methods for UGVs.

Main Methods:

• A novel enhanced pooling network was designed to adapt BRPN to objects of varying shapes.
• The BRPN loss function was improved for better learning of negative samples, optimized using the Grey Wolf Optimizer (GWO).
• A Genetic Algorithm-Support Vector Machine (GA-SVM) classifier was integrated to enhance classification capabilities.

Main Results:

• The BRPN demonstrated adaptability to diverse object shapes through the enhanced pooling network.
• Optimization of the BRPN loss function using GWO led to improved performance.
• The GA-SVM classifier significantly strengthened the overall classification capacity.
• The BRPN achieved excellent experimental results on PASCAL VOC 2007, VOC 2012, and KITTI datasets.

Conclusions:

• The developed Boosted RPN (BRPN) effectively enhances object detection for unmanned ground vehicles (UGVs).
• The proposed improvements, including an enhanced pooling network, optimized loss function, and GA-SVM classifier, contribute to superior performance.
• This deep learning-based object detection method shows significant promise for UGV environmental perception.