Problem and data
Define the task as detection, classification, counting, tracking or visual inspection. Camera, lens, distance, lighting and capture timing are assessed together. Success criteria require real samples and defined error categories.
Development and integration
Classical image processing or learned models are selected for the task. Video analysis, decision logic and API interfaces follow the operational workflow. Camera data, model output and business decisions are separate layers.
Evaluation and deployment
False acceptance and rejection are measured separately on test data. Lighting changes, occlusion and camera loss are assessed. Speed and hardware needs are verified on the actual stream; no unverified FPS or accuracy is stated here.