MBT3D: Deep learning based multi-object tracker for bumblebee 3D flight path estimation

Area of Science:

• Computer Vision
• Animal Behavior
• Robotics

Background:

• Tracking bumblebee flight paths in social groups is complex due to rapid movements, appearance variations, and individual similarities.
• Existing tracking algorithms often have limitations, including strict lighting requirements, high contrast needs, and susceptibility to occlusion.

Purpose of the Study:

• To develop and evaluate the Multi-Bees-Tracker (MBT3D), a Python framework for robust bumblebee tracking in social settings.
• To compare different deep learning detection architectures (YOLOv5, Faster R-CNN, RetinaNet) for optimizing bumblebee detection performance.
• To reconstruct three-dimensional (3D) flight paths of bumblebees using stereo camera data.

Main Methods:

• Implemented a deep association tracker (MBT3D) based on an ant tracking algorithm, incorporating an offline trained appearance descriptor and Kalman Filter.
• Trained detection models on 11,359 labeled bumblebee images and evaluated YOLOv5, Faster R-CNN, and RetinaNet performance.
• Utilized a validation dataset of 2,000 images to assess the tracker's accuracy (MOTA, MOTP) and identity switch rates.

Main Results:

• YOLOv5 achieved the highest Average Precision (AP) of 53.8% in bumblebee detection, outperforming Faster R-CNN (45.3%) and RetinaNet (38.4%).
• The MBT3D tracker, using Faster R-CNN detections, reached a Multiple Object Tracking Accuracy (MOTA) of 93.5% and MOTP of 75.6%.
• MBT3D demonstrated significantly lower identity switches and false positive rates compared to other state-of-the-art animal tracking algorithms.

Conclusions:

• The MBT3D framework provides reliable tracking of individual bumblebees within a group, minimizing identity switches.
• The study highlights the effectiveness of deep learning approaches for complex animal tracking tasks.
• MBT3D offers a valuable tool for reconstructing 3D bumblebee flight paths, advancing behavioral ecology research.