Image Segmentation Technique for Locating Automotive Parts on Belt Conveyors

Michael L. Baird

Abstract

This paper presents a robust, model-free image processing pipeline designed to locate automotive parts on dirty conveyor belts. By combining edge detection, automated thresholding, and morphological operations like gap and hole filling, the system generates clean silhouettes that allow robot arms to identify and manipulate parts in real-time industrial environments.

Introduction

In the mid-1970s, the "holy grail" for industrial automation was enabling robots to "see" and handle parts on a moving assembly line. The primary challenge was the inherent messiness of factory floors. Standard computer vision techniques of the era relied on simple thresholding—converting a grayscale image into black and white based on a single brightness value. However, on a dirty, well-used conveyor belt, the background is rarely uniform. Grease, scuffs, and dust create visual noise that causes simple thresholding to fail, resulting in lost parts or background merging. Michael L. Baird proposed a method to reliably extract clean object silhouettes (segmentation) without requiring pre-defined geometric models.

Methodology

The proposed system utilizes a unique sequence of image enhancement processes to transform noisy grayscale input into clean binary silhouettes.

  1. Edge Detection: A gradient operator (specifically the Sobel operator) highlights object boundaries.
  2. Smoothing: A 3x3 neighborhood average removes isolated noise and bridges minor gaps in boundaries.
  3. Automatic Thresholding: The system calculates a threshold $T$ by identifying the "right shoulder" of the edge-data histogram, specifically where the second derivative is maximized.
  4. Gap Filling: Neighborhood analysis bridges gaps up to 3 pixels using specific spatial configurations.
  5. Hole Filling & Noise Removal: 4-connectivity analysis fills hollow centers and removes small pixel clusters that do not correspond to actual parts.
The image processing pipeline: from edge enhancement (Fig. 3) and smoothing (Fig. 4) to histogram-based thresholding (Fig. 5) and the final binary silhouette (Fig. 6).

Key Insights

A core innovation is the "Right Shoulder" Heuristic for automatic threshold selection. By analyzing the histogram of edge data and finding the point of maximum second derivative on the curve's shoulder, the system adapts to varying lighting and belt conditions without manual tuning.

Furthermore, the approach is Model-Free. Unlike systems searching for specific shapes (e.g., bolts or gears), this method focuses on finding closed boundaries. This flexibility is crucial for dynamic factory environments where parts change frequently. Once a clean silhouette is obtained, the system calculates the "center of area" and "axis of minimum moment of inertia" for robotic manipulation.

Spatial configurations used for gap filling, showing the directional logic (North, South, East, West, etc.) required to close object boundaries.

Results and Discussion

The system was validated using digitized industrial images (244 x 256 pixels, 64 gray levels) of various automotive parts on dirty conveyor belts. It met production cycle time requirements, processing images in several seconds, and was successfully integrated with a prototype robot arm. The algorithm proved robust across different camera resolutions and hardware platforms, including mini-computers.

Limitations

The primary limitation is the "overlap problem": the system cannot distinguish between touching or overlapping parts, resulting in a single uninterpretable silhouette. Additionally, as it relies on silhouettes, it cannot resolve orientations for parts with identical external shapes but different internal features. Future work suggests incorporating simple models to flag or ignore silhouettes that do not match expected part dimensions.

Summary updated

BibTeX
@inproceedings{baird1977image,
  author = {Michael L. Baird},
  title = {{Image Segmentation Technique for Locating Automotive Parts on Belt Conveyors}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1977},
}
Figure