TL;DR
DeltaVisionX is an affordable industrial sorting system that bridges the gap between state-of-the-art computer vision (YOLOv11) and standard industrial controllers (PLC). By utilizing encoder-based synchronization rather than simple timers, the system achieves a high-precision sorting rate of 80 objects per minute on a variable-speed conveyor belt.
What Problem Does This Paper Address?
While industrial automation is a staple of developed markets, small and medium-sized enterprises (SMEs) in developing regions—specifically Lebanon, in this study—often find high-end robotic systems financially and technically inaccessible. These businesses frequently rely on manual labor for repetitive tasks like sorting, leading to inconsistent performance and high operational costs.
The technical challenge lies in the "integration gap." While you can easily find a powerful object detection model or a robust Programmable Logic Controller (PLC), making them talk to each other in real-time on a moving conveyor belt is difficult. Most low-cost setups use time-based triggers, which fail the moment a conveyor belt jitters or changes speed. DeltaVisionX tackles this by creating a synchronized pipeline that remains accurate even under realistic, non-deterministic industrial conditions.
Methods
The authors developed a three-part architecture: Image Acquisition, Detection/Classification, and Actuator Control.
- Vision Stack: Instead of expensive industrial cameras, the system uses a standard USB webcam and a laptop. It runs YOLOv11, the latest iteration of the "You Only Look Once" family, which is a "one-stage" detector optimized for high-speed inference.
- The Integration Strategy: This is the core novelty. The laptop (running Python) communicates with a Delta DVP-SE PLC via TCP/IP.
- Encoder-Based Synchronization: To solve the timing issue, the team mounted an encoder beneath the conveyor belt. When the vision system detects a chocolate bar, the PLC records the current "encoder pulse count." Because the distance between the camera and the pneumatic pistons is fixed, the PLC simply waits for the belt to move the required number of pulses before firing the piston. This makes the system immune to fluctuations in conveyor speed.
To handle multiple items on the belt at once, the authors implemented a queuing system using cyclic buffers (memory stacks) in the PLC. This ensures that if three "Snickers" bars pass the camera in quick succession, the PLC tracks all three and fires the corresponding piston three times at the exact right moments.
Datasets
The researchers built a custom dataset specifically for industrial sorting:
- Content: Images of various chocolate bars (Bounty, Snickers, Twix, etc.) captured at different angles, lighting conditions, and conveyor speeds.
- Labeling: Approximately 100 images per class were manually labeled using Roboflow.
- Augmentation: To improve robustness, the team used a custom script to apply rotations, noise addition, and "mosaic effects" (combining multiple images into one) to artificially expand the training data.
Key Insights and Ideas
- Integration > Algorithm: The authors emphasize that their contribution isn't a new AI model, but a scalable integration strategy. They proved that you don't need a $50,000 robot to achieve industrial-grade sorting if your synchronization logic is sound.
- Physical vs. Temporal Tracking: Shifting from "wait 2 seconds to fire" to "wait 500 encoder pulses to fire" is the "aha moment" that transforms a hobbyist project into a reliable industrial tool.
- Edge-Ready Architecture: By using a lightweight version of YOLOv11 (yolo11n.pt), they achieved real-time speeds on a standard laptop, proving that GPU-heavy server racks aren't always necessary for the factory floor.
Key Results
The system's performance was validated through both software metrics and hardware stress tests:
- Accuracy: The YOLOv11 model achieved a Mean Average Precision (mAP) of 0.97, indicating extremely high reliability in identifying the correct candy bars.
- Speed: Inference time was clocked at 100 ms per frame, well within the requirements for real-time operation.
- Throughput: The physical system successfully sorted 80 objects per minute.
- Reliability: Hardware testing confirmed that the pistons responded consistently across various conveyor speeds thanks to the encoder feedback.
Limitations and Future Work
Despite the success, the authors noted a few hurdles:
- Occlusion and Proximity: If two objects are stacked or touching, the system struggles to differentiate them. A suggested fix is a "mechanical bridge" at the start of the belt to force objects into a single-file line.
- False Positives: Transparent objects (like water bottles) were occasionally misclassified as candy due to color similarities. Future work involves expanding the "background" or "non-target" classes in the dataset.
- Scaling Up: The next phase of the project involves replacing the pneumatic pistons with a Delta robot (a high-speed parallel robot) and moving to EtherCAT communication for sub-millisecond control latency.
Summary updated