Improving the Robustness of Object Detection Through a Multi-Camera-Based Fusion Algorithm Using Fuzzy Logic.

Area of Science:

• Computer Vision
• Robotics
• Agricultural Technology

Background:

• Single RGB cameras with Convolutional Neural Networks (CNNs) can struggle to accurately localize objects within bounding boxes (BBs).
• Limitations in single-camera systems necessitate more robust object detection methods, especially in dynamic environments like agriculture.

Purpose of the Study:

• To develop and validate a multi-camera fusion algorithm that enhances the trustworthiness of object detection bounding boxes.
• To introduce a confidence scoring system for bounding boxes generated by CNNs in a multi-camera setup.

Main Methods:

• Implemented a two-camera system for detecting agricultural weeds.
• Utilized projective transformation to align image planes between cameras.
• Developed a Mamdani fuzzy logic system to fuse bounding box information and generate confidence scores based on Intersection over Union (IOU) overlap and object distance.

Main Results:

• The multi-camera fusion algorithm generated confidence scores ('high,' 'ok,' 'low') reflecting bounding box accuracy and position.
• IOU overlap values were calculated against ground truth across four distance scenarios.
• The confidence scores validated the hypothesis that the proposed system improves detection robustness.

Conclusions:

• The proposed multi-camera fusion approach significantly enhances the robustness of object detection systems.
• Fuzzy logic integration provides a reliable method for assessing the quality of bounding box detections.
• This system offers a promising solution for accurate object localization in agricultural applications.