Production CV Deployments at Collablens

Built and deployed computer vision systems for ITC freight monitoring and Smartivity defect detection

I joined Collablens in Dec 2022 as the founding senior research engineer and worked there until Jan 2024. Over 14 months, I built and deployed two production computer vision systems for industrial clients: a truck-loading monitoring system for ITC’s freight operations, and a defect detection system for Smartivity’s assembly line.

ITC Freight Damage Mitigation

A near-real-time monitoring system installed at ITC loading bays to detect rough handling of freight bags during truck loading.

  • Trained a YOLO detector to recognize impact patterns on superimposed video frames, a technique that compresses temporal motion into a single image so the model can learn what “rough handling” looks like visually
  • Tuned the pipeline to operate at 70% recall on rough-handling events while keeping false positives under 5%, which was the threshold ITC needed to act on alerts without alert fatigue
  • Built a parallel bag detection module that counted bags and classified bag types as they entered the loading zone, used for reconciliation against shipment manifests
  • Integrated three independent vision modules (impact detection, bag counting, bag classification) into a single deployed pipeline running in near real-time on factory floor hardware
Pose estimation locates the worker and the bag in frame, isolating the region used downstream by the motion-analysis stage.
Optical-flow pipeline. Left: raw clip flagged as a rough-handling violation. Center: cropped region zoomed onto the worker–bag interaction. Right: dense optical flow on that region — the high-magnitude motion vectors are the signal used to confirm the violation.
End-to-end pipeline running in near real-time on factory floor hardware.

Smartivity Defect Detection

A vision-based quality control system for Smartivity’s laser-cut MDF assembly kits, replacing manual inspection on the production line.

  • Built a classifier to detect missing parts in assembled units, comparing against a reference layout per SKU; reached 97% accuracy on the production dataset
  • Reduced per-unit inspection time from around 10 seconds (manual) to about 2 seconds (automated), enabling line-rate inspection
  • Designed a millimeter-precision deviation detection module using homography transformations to align captured images with reference templates before pixel-level comparison
  • Iterated on the physical rig (lighting, camera placement, enclosure) until detection was robust to factory floor conditions including variable ambient light and conveyor vibration

The system went through six iterations between the first sketch and the final production rig. The progression below walks through design, prototype, development, first on-site deployment, debugging under factory conditions, and the two final production modules.

Design. Initial dimensional sketches and component layout, drafted before fabrication.
Prototype. First physical build — PVC enclosure with mounted Raspberry Pi Camera Module 3.
Development. Bench-testing the imaging and inference pipeline before factory installation.
First Deployment. Rig installed at the Smartivity assembly line for the first on-site test.
Debugging. Iterating on lighting, camera placement, and conveyor alignment under factory floor conditions.
Final Deployments. Left: missing-parts detection running on the production line. Right: millimeter-precision defect detection comparing captured frames against the reference template.

Stack

Detection & CV: YOLO, OpenCV, homography-based image alignment Deployment: Raspberry Pi (edge inference), Python serving, AWS, GCP Hardware: custom PVC and aluminum-extrusion rig fabrication, Pi Camera Module 3