On the Feasibility of Estimating Fruits Weights Using Depth Sensors

Please cite:

• Mousavi, M. S. H., and VB Surya Prasath. "On the feasibility of estimating fruits weights using depth sensors." 4 th International Congress of Developing Agriculture, Natural Resources, Environment and Tourism of IranAt: Tabriz Islamic Art University In cooperation with Shiraz University and Yasouj University, Iran. 2019.
• Mousavi, Seyed Muhammad Hossein and Surya Prasath, V.B.,1398,On the Feasibility of Estimating Fruits Weights Using Depth Sensors,4th International Congress of Developing Agriculture, Natural Resources, Environment and Tourism of Iran,Tabriz,https://civilica.com/doc/971940
• DOI: http://dx.doi.org/10.13140/RG.2.2.33205.40162

📜 Paper Overview

Authors: Seyed Muhammad Hossein Mousavi, V. B. Surya Prasath
Conference: 4th International Congress of Developing Agriculture, Natural Resources, Environment, and Tourism of Iran (Feb 13–15, 2019)

This paper proposes a novel method to estimate fruit weights using RGB-D data captured by Microsoft Kinect V2. By leveraging depth information alongside RGB data, the method eliminates the need for expensive industrial weighing systems, making it a cost-effective and efficient solution for automated fruit weighing.

---

📝 Abstract

Automatic fruit weight estimation is essential in modern agriculture for improving efficiency and reducing human errors in weighing systems. Despite existing image processing-based systems, the use of consumer-grade depth sensors like Microsoft Kinect remains underexplored. This study evaluates the feasibility of combining RGB and depth data to estimate weights for Sweet Lemons, Sweet Peppers, and Tomatoes.

Key features:

• Dataset: RGB-D data of 50 samples for each fruit, captured using Kinect V2.
• Metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE) were evaluated at distances of 0.8m, 1.0m, and 1.3m.
• Results: Achieved a recognition accuracy of 94.7% using the proposed method.

---

🚀 Methodology

1. Data Acquisition:

   • RGB-D data was captured using Microsoft Kinect V2.
   • 150 fruit samples (50 each for Sweet Lemons, Sweet Peppers, and Tomatoes) were imaged at three distances: 0.8m, 1.0m, and 1.3m.

2. Preprocessing:

   • Noise reduction applied to RGB and depth images.
   • Depth normalization performed to adjust for sensor-object distances.

3. Segmentation:

   • Thresholding used to isolate fruit objects from the background in RGB and depth images.

4. Weight Estimation:

   • Formula: [ \text{Weight} = \frac{\alpha \times \beta \times \delta}{\theta} ]
     • (\alpha): Average dimensions of the segmented RGB image.
     • (\beta): Distance from the depth sensor (in meters).
     • (\delta): Average weight of the fruit (based on dataset statistics).
     • (\theta): Normalization constant (set to 100).

5. Evaluation:

   • Metrics: MAE, MSE, RMSE, Mean Error (ME), Standard Deviation Error (STDE).

---

📊 Results

Quantitative Performance:

Metric	Sweet Lemons	Sweet Peppers	Tomatoes	All Samples
MAE	5.4	4.5	6.2	5.3
MSE	7.58	6.31	8.07	7.32
RMSE	2.75	2.51	2.84	2.70

• Best Performing Fruit: Sweet Peppers (lowest MAE and RMSE).
• Overall Recognition Accuracy: 94.7%.

---

📦 Dataset

A custom dataset was created using Microsoft Kinect V2:

• Fruits: Sweet Lemons, Sweet Peppers, and Tomatoes.
• Samples: 50 for each fruit (total: 150 samples).
• Features: RGB and depth data at three distances (0.8m, 1.0m, 1.3m).

---

💡 Future Work

1. Extend the method to other fruit types and shapes.
2. Explore multi-directional depth sensors for improved accuracy.
3. Implement real-time fruit weighing on moving robotic systems.

---