Mango Fruit Load Estimation Using a Video Based MangoYOLO-Kalman Filter-Hungarian Algorithm Method.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Pre-harvest fruit yield estimation is crucial for efficient harvesting and resource allocation in orchards.
• Traditional machine vision methods using dual-view imaging often underestimate yield due to fruit being hidden within tree canopies.

Purpose of the Study:

• To develop and evaluate an advanced on-tree mango fruit detection, tracking, and counting system using video analysis.
• To improve the accuracy of fruit yield estimation compared to existing dual-view methods.

Main Methods:

• Utilized a deep learning algorithm (MangoYOLO) for fruit detection in video frames.
• Employed the Hungarian algorithm for correlating fruit across frames and a Kalman filter with a 'borrow' concept for robust tracking and occlusion handling.
• Captured 10 fps video data from a moving platform at 5 km/h.

Main Results:

• The proposed video tracking system achieved a bias-corrected Root Mean Square Error (RMSE) of 18.0 fruit/tree, detecting 62% of the total harvest.
• The dual-view method detected only 40% of the harvest with a higher RMSE of 21.7 fruit/tree.
• Initial tests showed a low over-count error of approximately 2.6%.

Conclusions:

• The video tracking system significantly outperforms the dual-view imaging system for mango orchard fruit counting, especially in dense canopies.
• This advanced method provides a more accurate and reliable approach to pre-harvest fruit yield estimation.
• Recommended for practical application in mango orchard management and yield prediction.