Automated Pest Counting in Water Traps through Active Robotic Stirring for Occlusion Handling

This paper proposes an automated pest counting system for water traps that utilizes a robotic arm with adaptive-speed stirring and a confidence-driven closed-loop control mechanism to effectively mitigate occlusion, significantly reducing counting errors and execution time compared to static image methods and constant-speed stirring.

The Gist

Imagine you are trying to count a swarm of tiny, mischievous ants floating in a bowl of yellow water. The problem? They love to huddle together, hide under each other, and form little clumps. If you just take a single photo of the bowl, you'll likely miss many of them because they are stacked on top of one another. It's like trying to count people in a crowded mosh pit from a single, frozen photo—you'll definitely underestimate the crowd.

This paper presents a clever solution: Don't just look; stir!

Here is the story of how the researchers solved this problem, explained simply:

1. The Problem: The "Mosh Pit" Effect

In agriculture, farmers use "water traps" (basically bowls of yellow liquid) to catch pests so they can monitor crop health. Traditionally, humans have to take these bowls to a lab and count the bugs under a microscope. It's slow, boring, and prone to human error.

Computer programs (AI) can try to count them from photos, but they fail when the bugs are crowded. If three bugs are stacked, the computer sees one big blob, not three bugs. This is called occlusion.

2. The Solution: The Robotic Bartender

Instead of just taking a photo, the researchers built a robot arm with a stick attached to it. Think of this robot as a robotic bartender who knows exactly how to mix a drink to separate the ingredients.

The robot dips the stick into the water trap and stirs the water. This movement scatters the bugs, breaking up the clumps and revealing the ones that were hiding underneath. But here's the catch: if you stir too hard, the water gets too wavy and blurry to see anything. If you stir too gently, the bugs stay stuck together.

3. Finding the Best Dance Move (Stirring Patterns)

The team asked: What is the best way to stir? They didn't just guess; they tested six different "dance moves" for the robot stick:

• The Circle: Going round and round (the traditional way).
• The Square & Triangle: Moving in geometric shapes.
• The Spiral: Winding inward like a snail shell.
• The Random Lines: Zipping back and forth unpredictably.
• The "Four Circles": A unique pattern where the robot draws four small circles connected by lines.

The Winner: Surprisingly, the traditional "Circle" was the worst performer. The winner was the "Four Circles" pattern. It was like the robot was doing a specific, intricate footwork that scattered the bugs perfectly without making the water too messy.

4. The "Smart" Stirrer (Adaptive Speed)

Once they found the best pattern, they realized the robot shouldn't just stir at a fixed speed. It needs to be smart.

Imagine you are trying to untangle a knot of headphones. You pull gently, check the knot, pull a bit harder, check again, and then stop when it's loose.

• The Old Way: The robot would stir at a constant speed until a timer ran out. It might keep stirring even after the bugs were separated, making the water messy and ruining the photo.
• The New Way (Adaptive): The robot has a "brain" that looks at the photos in real-time. It asks: "Are the bugs getting easier to count?"
  • If the count is getting clearer, the robot speeds up to finish the job faster.
  • If the water is getting too wavy and the count is getting confusing, the robot slows down.
  • If the count stops changing (meaning the bugs are fully separated), the robot stops immediately.

This "smart" approach saved 44% of the time compared to the dumb, constant-speed robot.

5. The Final Count

The robot doesn't just take one photo at the end. It takes a whole video sequence while stirring. It counts the bugs in every frame, but it trusts the clear frames more than the blurry ones. It combines all these counts into one final, highly accurate number.

The Result

• Accuracy: By stirring, the system reduced counting errors by a huge margin, especially when the bugs were packed tight (high density).
• Speed: The "smart" stirring was much faster than stirring at a fixed speed.
• Reliability: It worked consistently across different amounts of bugs, from a few to a swarm.

In a Nutshell

This paper teaches us that sometimes, to see the whole picture, you have to shake things up. By using a robot to intelligently stir a water trap, the researchers turned a blurry, confusing mess of hidden bugs into a clear, countable crowd. It's a perfect example of using active perception—not just waiting to see, but interacting with the world to get a better view.

Technical Summary

1. Problem Statement

Accurate monitoring of pest populations is critical for preventing crop damage and economic loss. Traditional methods involve manual counting under microscopes, which is labor-intensive. While image-based automatic counting exists, it relies on single static images, leading to significant inaccuracies due to occlusion (pests overlapping or hiding behind one another) in water traps.

Existing automated stirring solutions have two major limitations:

1. Lack of Pattern Optimization: Most rely on standard circular stirring, ignoring how different geometric patterns affect pest redistribution.
2. Lack of Adaptive Control: They typically use constant-speed, open-loop control. Adjusting speed based on the real-time state of the liquid (fluid dynamics) is challenging, leading to either under-stirring (missing occluded pests) or over-stirring (creating turbulence that degrades image quality).

2. Methodology

The authors propose a closed-loop system integrating robotic manipulation with computer vision. The system consists of four main components:

A. Robotic Stirring System

• Hardware: A Franka robotic arm equipped with a wooden stirring stick manipulates pests in a yellow water trap.
• Vision: A top-down camera captures images during the stirring process.
• Control: A PC running ROS 2 Humble coordinates motion planning (MoveIt 2) and image acquisition.

B. Pest Detection and Counting

• Model: An improved YOLOv5 architecture is used for detection, featuring:
  • ODConv: Dynamically adjusts convolutional kernels for better feature extraction in dynamic scenes.
  • CoT3 Block: Combines local CNN features with global Transformer-based context.
  • Soft-NMS: Replaces standard Non-Maximum Suppression to better handle densely packed pests.
• Preprocessing: GroundingDINO is used to crop the specific water trap region, which is then resized to a square for the detector.

C. Counting Confidence Assessment

• Mechanism: A polynomial regression model estimates the reliability (confidence) of a count for each frame.
• Input Features:
  • Detection metrics (mean confidence, predicted count).
  • Image quality (NIQE score, entropy, gradient magnitude).
  • Pest distribution uniformity (DBSCAN clustering).
• Output: A confidence score ($C$) based on the Jaccard index against ground truth.

D. Adaptive-Speed Closed-Loop Control

• Logic: The system adjusts stirring speed ($S$) based on the average change rate of counting confidence ($\Delta C_T$) between consecutive frames.
  • If $|\Delta C_T| > \text{Threshold}$: The system adjusts speed ($S_T = S_{T-1} + \Delta C_T$). A positive $\Delta C$ suggests occlusion is being resolved (speed up); a negative $\Delta C$ suggests excessive turbulence (slow down).
  • If $|\Delta C_T| \le \text{Threshold}$: Stirring terminates, assuming pests are sufficiently redistributed.
• Final Count: The final count is a weighted sum of counts from all frames in the sequence, where weights are derived from the softmax probability of the confidence scores.

E. Stirring Pattern Evaluation

Six distinct patterns were designed and tested to find the optimal trajectory:

1. Circle (Reference)
2. Square
3. Triangle
4. Spiral
5. Four Circles
6. Random Lines

3. Key Contributions

1. Active Robotic Stirring System: Development of an automated system that physically redistributes pests to reveal occluded individuals, moving beyond static image analysis.
2. Pattern Optimization: A systematic evaluation of six stirring patterns, identifying that non-traditional patterns (specifically "Four Circles") outperform the standard circular motion.
3. Heuristic Adaptive Control: Introduction of a counting confidence-driven closed-loop control mechanism. Unlike previous open-loop methods, this system dynamically adjusts stirring speed based on visual feedback, balancing pest redistribution with image stability.
4. Robust Counting Framework: A method to aggregate multi-frame data using confidence weighting, significantly reducing errors caused by transient occlusions.

4. Experimental Results

Experiments were conducted across three pest density scenarios (Low, Medium, High) with 20 trials each.

A. Optimal Stirring Pattern

• Winner: The "Four Circles" pattern achieved the best overall performance.
  • Lowest Mean Absolute Error (MAE1): 4.384
  • Highest Mean Counting Confidence (MCC): 0.721
• Comparison: The commonly used "Circle" pattern performed the worst (MAE1: 4.973). "Random Lines" performed best only in medium density.

B. Adaptive-Speed vs. Constant-Speed

• Efficiency: Adaptive-speed stirring reduced task execution time by 39.2% to 44.7% compared to constant speed.
• Stability: The standard deviation of execution time decreased by 52.9% to 77.8%, indicating much more consistent performance.
• Accuracy: Both methods achieved similar final counting errors (MAE1), but adaptive stirring maintained higher counting confidence (MCC) by avoiding unnecessary turbulence.

C. Comparison with Static Image Counting

• The proposed robotic stirring method significantly outperformed single static image counting, especially in high-density scenarios.
• High-Density Improvement: Reduced Mean Absolute Counting Error (MAE2) by 3.428 compared to the static method.
• Limitation: While improved, errors remain in high-density scenarios due to physical adhesion of pests and fluid dynamics causing re-occlusion.

5. Significance and Future Work

• Significance: This work bridges the gap between robotic manipulation and computer vision for agricultural monitoring. It demonstrates that active perception (stirring) combined with adaptive control is essential for solving occlusion problems in dynamic liquid environments.
• Limitations:
  • Only six patterns were tested; more complex shapes (e.g., figure-eight) may exist.
  • The confidence estimation pipeline has a latency of 1–2 seconds, preventing true real-time control for very rapid pest movements.
• Future Directions: Exploring a wider range of stirring patterns and optimizing the confidence assessment pipeline for lower latency to enable faster, more responsive closed-loop control.