IVS / Edge AI Research Brief

Signals (10)

1. 1

   Flock goes on the record: Traffic Analytics is a separate SKU, 5-year aggregates, “we do not sell the data”

   Key point

   After Racine/Fond du Lac/Sheboygan, Flock told FOX6 it does not sell camera data. Two WI failure modes: traffic data and out-of-state FOIA of search logs. Vendor blog: Traffic Analytics (shipped 2025) is not the LPR platform — aggregate volume, vehicle mix, registration-state share, and 15-minute group speed stats; no plates / no individual vehicles in the analytics product; LPR records still delete on the agency’s retention clock; aggregates may be kept 5 years. Sold to LE, city/county/state DOTs, and municipalities. Agency that owns the cameras owns the derived analytics; a DOT does not get it without that agency’s permission. FOX6: Flock stripped plates and officer names from the network audit log after “nearly 200” LE leaders said public audit dumps were outing undercover officers. Sheboygan: 15 cameras bagged after a sales email offered Engineering 5 years of speed/volume from Sheboygan’s own cameras; Common Council votes today. Do not treat the vote as closed.

   Why it matters

   Yesterday’s brief named the SKU; today the vendor defined it. RFP language is now copy-paste: LPR SaaS ≠ Traffic Analytics, retention parity, no secondary commercial dataset, who is the records custodian when another state’s FOIA hits the audit log. “We don’t sell data” and “we sell a 5-year aggregate product to DOTs” can both be true — write the contract as if they are.

   Source

   FOX6; Flock Safety blog (Sept 5)
2. 2

   Kenosha: terminate Flock a year early, keep ALPR — 10-year Axon with 31 Outpost poles

   Key point

   Chief Patrick Patton: cancel Flock one year early over undisclosed tracking, data sharing, and security. No local officer-abuse finding. City already signed a 10-year, $14.6M Axon bundle: 31 Outpost fixed LPR cameras, 100 mobile plate readers, 200 BWCs, 227 Tasers, 3 drones, evidence cloud. Patton’s trade: greater control over who queries — no nationwide search; share Wisconsin and possibly Lake County, IL. Process started January 2026. Mayor David Bogdala joint statement: ALPR stays, with supervisor/search rules. Council is the formal Flock-exit vote today; the Axon swap is already signed.

   Why it matters

   This is Racine’s slogan as a 10-year PO: keep plates, change vendor, shrink the graph. Integrators bidding “post-Flock” WI/IL work should quote Outpost + in-state share, not “no ALPR.” Nationwide lookup is now a negotiated feature, not a default.

   Source

   FOX6; City of Kenosha; Axon contract PDF
3. 3

   Harris County Judge Hidalgo calls a special Commissioners Court — renewal passed without the promised oversight

   Key point

   Hidalgo: May 28 Business Court (short, “non-controversial,” limited discussion) renewed Flock LPR + sound through June 5, 2027. Appendix purchasing figure $2,868,975 (her post body said “$2M”; some local TV used $868,975 — use the appendix). Vote 4–0, Hidalgo abstained. Court promised to add privacy/oversight later; as of September the amendment is not in. She wants a public special meeting covering: timeline, which precincts appear in the contracts, who can FOIA the data, which agencies have access, how many ALPRs and who administers them (SO “in the dozens”; “thousands” of ALPRs in the county including MUDs/HOAs/private), and a pause until oversight is written in. Cites Deer Park / Katy / Pasadena / Baytown misuse and Galleria vandalism. County can still cancel/amend on 30 days’ notice.

   Why it matters

   Largest remaining Texas Flock footprint is a governance fight, not a bag-the-poles fight. Integrators: “renew now, privacy rider later” is now an exhibit. Map county-paid vs MUD/HOA/private cameras — Hidalgo is asking the court to publish that split.

   Source

   Lina Hidalgo Substack (primary); KPRC; ABC13
4. 4

   Arizona: 11 cities off Flock; Queen Creek keeps the cameras with a class-1 misdemeanor for knowing misuse

   Key point

   Cottonwood and Payson cancellations take Arizona Flock exits to 11 — KJZZ, citing the Institute for Justice tracker, says that is the third-highest cancellation rate in the country. Queen Creek voted unanimously for Ordinance 894-26: knowing misuse (personal use, stalking/harassment, selling data) is a class 1 misdemeanor under town code. Chief Randy Brice: no Arizona state crime for Flock misuse because the data sits outside CJIS. 30-day retain; contracts must bar third-party sale; share today is local PDs only, no federal / no immigration / no reproductive — but the ordinance’s sharing line is “verified and authorized recipient,” which public comment called a permission slip. 47 ALPRs in service, 15 more in testing; Brice claims assist on >3,000 incidents since 2025. Annual public audit required (staff described more frequent internal review).

   Why it matters

   The keep-path is now a municipal criminal ordinance, not a vendor ToS. AZ has no state ALPR statute; towns are writing the crime themselves. Copy Queen Creek’s misuse definition into RFPs, and do not treat “we don’t share with ICE” as binding unless it is in the ordinance and the partner-agency graph.

   Source

   KJZZ; Queen Creek ordinance PDF
5. 5

   Milwaukee keeps Flock — public transparency portal plus supervisor-only search

   Key point

   MPD launched Flock’s public dashboard: camera count, 30-day queries, hotlist reads, alert types, offense types, retention/sharing settings, policy summaries. Mayor Cavalier Johnson (Aug 31 package): supervisors only (no officers/detectives), limit to Part I / critical missing / drug trafficking / abduction, no out-of-jurisdiction share, semiannual independent audits via the Fire and Police Commission. Milwaukee County SO launched a matching portal and cut retention to 7 days, in-state only. Keep, don’t pull.

   Why it matters

   The other fork from Kenosha: same vendor, thinner graph, public metrics. Expect “transparency portal + supervisor-only + Part I” as the keep slide in every council packet this month. If you sell a competing LPR, you need a public dashboard on day one or you lose to Flock’s portal even when the PD is under a stalking scandal.

   Source

   FOX6; CBS 58; portal
6. 6

   QNX SDP 8.0 adds Hailo-8 — RTOS + NPU for boxes, not a new camera SKU

   Key point

   BlackBerry QNX enabled Hailo-8 on SDP 8.0 for automotive, industrial, robotics, medical, and smart infrastructure. Pi 5 + Hailo-8 vs RT Linux: up to 14× performance consistency, 2.6× tighter latency distribution, +4.1% throughput, −3.9% average latency. Whitepaper: QNX for Deterministic Physical AI Workloads. Hailo CRO Max Glover: broader embedded AI without giving up determinism. This is accelerator support on an RTOS, not a new dome camera.

   Why it matters

   Box and appliance builders can put a swappable Hailo-8 under QNX instead of a locked SoC camera. Accelerated compute beats locked “AI cameras”: swap models vs a 36–84 month hardware refresh.

   Source

   QNX press release (Aug 27); Automotive World; IT Brief US recirculation (2026-09-07)
7. 7

   NVIDIA: frontier reasoning models on Jetson — NVFP4 + speculative decode up to 6.28× vs BF16

   Key point

   Jetson blog: compact 2026 open models now run multi-step reasoning locally. Worked examples: Nemotron 3.5 Lightning (30B MoE, 3B active/token) and dense Qwen3.8-27B. NVFP4 plus speculative decoding (DSpark on Lightning, DFlash2 on Qwen) hits up to 6.28× decode throughput vs BF16 on Jetson. Gemma 4 E4B for Orin Nano; Lightning/Qwen for AGX Orin and AGX Thor. vLLM 0.28 recipes, JetPack 7.2. Use case they name: on-site anomaly loops and agents that stay up when the WAN is down.

   Why it matters

   The IVA box is becoming an on-prem reasoning host, not just a YOLO appliance. Spec it as Jetson/Thor + quantized checkpoint, not a 4-year locked analytics camera. Accelerated compute beats locked “AI cameras”: swap models vs a 36–84 month hardware refresh.

   Source

   NVIDIA Technical Blog
8. 8

   USI ships an edge AI smart camera for factory inspection — ODM, on-device, production-proven in its own plants

   Key point

   Taiwan ODM USI launched a packaged AI Smart Camera: high-res / low-lux module + low-power edge compute + proprietary vision software, no-code/low-code train-to-deploy. Claims: object/defect/OCR, key-part tracking, assembly verify, behavior analysis; industrial I/O (Ethernet, PoE, HDMI, MicroSD). Already running in USI’s own factories as the reference. Director Justin Chang: customers want an end-to-end platform, not a bare camera.

   Why it matters

   48-hour SKU, not a city ALPR. Plant/warehouse/MSPs get another locked ODM camera with the model inside the housing. Prefer the same job on a swappable NPU box (Hailo/Jetson/Dragonwing) unless the line truly needs a sealed imager. Accelerated compute beats locked “AI cameras”: swap models vs a 36–84 month hardware refresh.

   Source

   PR Newswire / Manila Times
9. 9

   arXiv: VLM as a tracker tuner — no target labels, recovers most of the MOT17→MOT20 drop

   Key point

   Cross-domain detect-to-track adaptation without target-domain labels. A VLM inspects rendered tracking, names visual failure modes, and iterates parameter updates. MOT17→MOT20: source oracle HOTA 0.357 → 0.267 (−0.090); VLM tuner recovers 67.8% of that lost headroom (within 0.029 of the target ceiling); up to 86.7% on the densest sequence. It declines to change configs when it sees no visual failure (protects easy transfers). Weaker on MOT17→DanceTrack, where the source oracle is already near-optimal. Works when the shift is in exposed detection parameters, not when the tracker is already maxed.

   Why it matters

   Site density changes (lobby vs. stadium, day vs. event) usually mean a labeled retune. This is a path to label-free tracker retune with a VLM in the loop — useful on Jetson-class boxes if you keep the tracker’s knobs exposed instead of burying them in a locked camera.

   Source

   arXiv:2609.05239 (Davila, Balakrishnan, Cochran)
10. 10

    arXiv: weakly supervised video anomaly detection at event level, not fixed top-k snippets

    Key point

    WSVAD/MIL usually scores snippets with rigid top-k heuristics that break when events are short, long, or bursty. Temporal Refinement Module (dynamic positional encoding + class token) plus Event Segmentation Module (temporal discontinuities → event-level reps) plus adaptive similarity fusion instead of fixed top-k. Authors report SOTA on two standard WSVAD benchmarks.

    Why it matters

    This is the VAD job MSPs actually have: loitering vs. a 2-second fight vs. a 20-minute spill. Event-level aggregation is the design to ask vendors for; clip/top-k is what ships in most NVR “AI packs.”

    Source

    arXiv:2609.05066 (Li, Xiao; accepted PRCV 2026)