A Comparison Between Single-Stage and Two-Stage 3D Tracking Algorithms for Greenhouse Robotics

Area of Science:

• Robotics and Automation
• Computer Vision
• Agricultural Technology

Background:

• Automation in the agro-food industry requires accurate 3D object detection and localization for robotic operations.
• Occlusions pose a significant challenge to 3D object detection and tracking in robotic applications.
• Multi-view perception and multi-object tracking (MOT) are crucial for overcoming occlusions by associating objects across different viewpoints.

Purpose of the Study:

• To compare the performance of a 3D two-stage MOT algorithm (3D-SORT) against a 3D single-stage MOT algorithm (MOT-DETR).
• To evaluate tracking accuracy in varying levels of complexity relevant to robotic arm movements in a tomato greenhouse.
• To determine the effectiveness of single-stage versus two-stage MOT methods in handling occluded objects during robotic operations.

Main Methods:

• Comparison of 3D-SORT (two-stage MOT) and MOT-DETR (single-stage MOT) algorithms.
• Experiments conducted using three distinct sequences representing simple to complex robot arm motions in a tomato greenhouse environment.
• Evaluation of tracking accuracy based on object visibility and occlusion levels across multiple viewpoints.

Main Results:

• The single-stage MOT-DETR algorithm demonstrated consistently higher tracking accuracy compared to the two-stage 3D-SORT algorithm.
• Superior performance of the single-stage method was particularly evident in more challenging sequences involving full occlusions or extended periods of non-visibility.
• Tracking accuracy improvements were observed across all tested sequences, highlighting the robustness of single-stage approaches.

Conclusions:

• Single-stage 3D MOT algorithms are more effective than two-stage methods for robotic automation in complex agricultural environments like greenhouses.
• MOT-DETR provides superior tracking accuracy, especially when dealing with significant object occlusions, which are common in real-world agricultural tasks.
• The findings support the adoption of single-stage MOT methods for enhancing the reliability and efficiency of automated agricultural systems.