Site class dictates camera mount height, expected subjects, lighting infrastructure, and power envelope — a rooftop urban macro and an off-grid solar monopole have almost no shared assumptions. Rural lattice sites dominate copper and battery theft losses in the US, while urban small cells fail more often on vandalism and tamper. Scoping to the wrong mix produces a model that works everywhere on paper and nowhere in production.

Detection quality is a per-class problem — the model that catches fence breaches at 98% recall may be blind to in-progress copper theft inside a cabinet. Copper theft alone costs US telecom operators roughly $1B annually per industry reports, and climb-detection has distinct safety-regulatory implications under OSHA 29 CFR 1926 Subpart M. A comprehensive threat inventory anchors the labeling schema, the evaluation set, and the escalation playbook downstream.

Every second between a detected climb and an escalated alert directly raises the probability of a successful theft or a human injury. Cloud-routed analytics introduce a latency floor set by backhaul RTT and cloud queue depth, which is fundamentally incompatible with a sub-second safety-critical class. Latency tiering by threat class is almost always the right architecture — climb and weapon detection deserve a tighter budget than cabinet tamper.

Edge CV is a power problem before it is an accuracy problem. Sustained inference on 4–8 camera streams can draw 40–150W depending on model and accelerator — and on a battery-backed solar site that headroom is not available without sacrificing uptime on the radio load. Specifying the envelope up front drives every downstream decision from accelerator selection (Jetson class) to how many streams per node are realistic.

Per-site camera count is the first-order driver of edge-compute sizing — a 2-camera small cell and an 8-camera compound need different accelerator SKUs even with the same model. Typical macro towers converge on 2–4 cameras covering perimeter, tower base, cabinet, and access gate. Overspecifying camera count wastes power and capex; underspecifying leaves blind spots on the highest-loss sites.

Tower site video reveals equipment inventory, cable routing, access patterns, and authorized personnel — precisely the intelligence a sophisticated thief or threat actor needs. Continuous cloud streaming both creates that exposure and costs roughly $300–$800 per site per year in backhaul, which at 10,000+ sites is a material OpEx line. Defining the egress envelope before architecture selection locks the correct trust boundary in place.

Running face recognition or gait-based identification on tower-site video puts the program inside EU AI Act Annex III as a high-risk system and inside Illinois BIPA with per-capture statutory damages — a very different compliance posture than object and event detection. The cleanest architectural move is to design the model to classify events and objects, not to identify individuals, and to document that scoping choice. BIPA and CUBI have both produced material settlements against employers running biometric systems on workers without written consent.

Climb detection at an unauthorized hour overlaps OSHA 1926 Subpart M — if a worker is detected climbing without fall-protection indication, that is both a security event and a safety event, and the escalation path is different. FCC 47 CFR Part 17 mandates obstruction-light monitoring on many towers, and CV can be an inexpensive secondary monitor for lighting failures that also have FCC Part 4 reporting consequences. Mapping to ASIS and SIA performance standards keeps the deployment defensible against insurers and in post-incident review.