Jetson-Based Large Foreign-Object Detection System

The contractor will develop and deploy a real-time NVIDIA Jetson-based vision system to detect large foreign objects on a crushed-glass conveyor before they reach the crusher. Model: - Use a real-time YOLO instance-segmentation model, such as Ultralytics YOLO-seg. - Train the model with one class only: `large_foreign_object`. - Cardboard, wood, bags, bottles, plastic, clothing, and other crusher-risk items will use the same class. - PatchCore or other anomaly-detection models are not required unless testing demonstrates a clear performance benefit. Dataset: - The customer will provide approximately 2,500 images. - Review, clean, and label all relevant foreign objects as `large_foreign_object`. - Remove duplicates and near-duplicate frames where appropriate. - Split the dataset by conveyor run or date, not simply by adjacent frames: - 70% training - 15% validation - 15% final testing - Include normal crushed-glass images with no large foreign object. - Include variation in object type, size, orientation, colour, lighting, loading density, conveyor position, and partial burial. - Report performance separately for known object types, previously unseen object types, partly buried objects, and objects near the minimum dangerous size. Hardware and deployment: - Deploy and optimise the system for the agreed NVIDIA Jetson device. - Use TensorRT or another Jetson-compatible accelerated runtime where appropriate. - Demonstrate real-time operation at the conveyor’s 30 fps video rate. - Configure automatic startup and continuous operation without manual intervention. System requirements: - Detect large foreign objects without identifying their material or object type. - Ignore normal crushed glass, small fragments, reflections, and ordinary material variation. - Apply a configurable minimum object-size threshold based on crusher-blockage risk. - Track detections across multiple frames. - Generate one alert per physical object. - Provide the object location, segmentation mask, confidence score, timestamp, and captured image. - Output an alert through the agreed HTTP endpoint Deliverables: - Trained YOLO-seg model. - Jetson-deployed inference application. - Optimised Jetson model, such as TensorRT. - Adjustable confidence and minimum-size thresholds. - Alert/output integration. - Automatic startup and continuous-operation configuration. - Test report covering detection rate, false alerts, unseen-object performance, processing speed, and known limitations. - Basic operating instructions. Acceptance criteria: - Detects agreed crusher-risk objects at the required conveyor speed. - Runs continuously on the