IVS Briefs
IVS / Edge AI Research Brief
Overall assessment
Signal level
High on two overnight Flock kills (Newburgh unwind, South Lyon cover-tonight), SF’s 30-day keep, and a Florida statewide ROW deadline today.
Notable trends
- Kill path is now cover/dark in hours (South Lyon, Florida) plus legal unwind (Newburgh).
- Big cities are choosing keep-with-CAD-gates (SF 30-day / off-platform).
- Visual AI agents moved from posters to a Metropolis 3.3 compose skill and two long-video memory papers in the same 48h.
Gaps
Dallas meeting has no result primary yet. No new Hailo / DeepX / Qualcomm / Jetson module SKU in 48h (Edge is VSS software on Blackwell GPUs). 404 Media HIDTA is paywalled. Newburgh roll-call unpublished.
Mix counts
LPR 2 / VMS 1 / Edge 1 / Policy 2 / Lab 2. LPR+Policy city/contract = 4/8 = 50%.
Now vs On deck
Now 6 (items 1–6) · On deck 2 (VideoLoop, Video-RSI).
Signals (8)
- 1
Key point
After months of delayed votes, Newburgh City Council passed a resolution Monday night to discontinue routine use of Flock cameras within 14 days and take all legal steps to dissolve the contracts. City audit logs showed thousands of out-of-jurisdiction searches of Newburgh’s data, including a 2025 Johnson County, TX search for a “female” who “had an abortion.” The contract began Aug 6, 2024 and was renewed Jan 1, 2026. WAMC does not publish the roll-call tally.
Why it matters
This is a kill-and-unwind, not a funding pause. Bid named sharing lists + case-number logs or lose Hudson Valley rooms that have already read the Texas search.
- 2
Key point
Council voted 5–2 to notify Flock it will not renew, ordered PD to stop accessing, searching, or sharing the data (except as required by law), and directed officers to cover the two cameras on 10 Mile Road until Flock pulls them. PD posted at 11:45 p.m. that the lenses were covered and no longer capturing. The pair went in November 2025. South Lyon joins Ferndale, Westland, and Independence Township.
Why it matters
Small-city playbook is now cover tonight, unwind later. Quote immediate dark + no-share language, not a 60-day removal clause.
- 3
Key point
Mayor Daniel Lurie imposed administrative rules (no Board vote): SFPD may query ALPR data for 30 days after capture, down from 12 months. After 30 days the data leaves Flock’s platform, so only the department holds it. Every search must carry a CAD or incident number. Lurie will partner with Assemblymember Matt Haney and DA Brooke Jenkins on criminal penalties for misuse. ACLU-NC: internal supervision is not a warrant; officers have used junk justifications (“LMAO,” “IDK”). Flock: since 2024, SF monthly vehicle theft −55%, overall theft −45%, property crime −44.4%, 3–4 arrests/day. Board renewal is still February 2027.
Why it matters
Guardrail-keep is the big-city path. Bid 30-day vendor retain + CAD-gated search + off-platform archive, and expect a criminal-misuse statute in the next California packet.
Source
CBS News Bay Area - 4
Florida FDOT deadline is today — state-road ALPRs must be down; JSO already dark, Port St. Lucie pulls 18
Key point
FDOT revoked state-ROW ALPR permits ~30 days ago; Wednesday is the removal deadline. Jacksonville Sheriff T.K. Waters: units are covered and not in use; hardware owners (not JSO) pull the poles; JSO discontinued LPR countywide on Aug 31. Nassau, Clay, St. Johns, Jacksonville Beach gave little or no comment. Port St. Lucie: vendor is to take down 18 cameras (9 county ROW + 9 state ROW) by Sept 30; remaining use is limited to an immediate threat to life with Chief Leo Niemczyk’s authorization. St. Lucie County already voted 3–2 to stop new ALPR permits on county ROW.
Why it matters
State DOT permit is now a kill switch. Quote ROW owner + vendor-remove SLA, and do not bid Flock/Axon fixed ALPRs onto Florida state roads until FDOT relents.
- 5
Key point
Metropolis Video Search and Summarization 3.3 ships two cost cuts. Build Vision Agent (vss-build-vision-ai) composes alerting + search + summarization from a prompt on four validated profiles (base / alerts / lvs / search); a bottling-line demo reached a live stack in <30 min on a two-GPU RTX PRO 6000 Blackwell, with FP8 Cosmos 3 Nano sharing the detector GPU. Adaptive Efficient Video Sampling prunes unchanged patches and batches VLM work around events. On Cosmos 3 Super FP8: alert contextualization 1,021 → 844 ms (−17%), concurrent real-time VLM streams 13 → 19 (+46%), a 60-minute summary in ~half the time with 80% fewer VLM input tokens. Live walkthrough Oct 1, 9 a.m. PT. GitHub: NVIDIA-AI-Blueprints/video-search-and-summarization.
Why it matters
Quote Adaptive EVS + one-prompt compose on the next VLM-video RFP. Accelerated compute beats locked “AI cameras”: swap models vs a 36–84 month hardware refresh.
Source
NVIDIA Technical Blog - 6
Key point
Hanwha Vision’s XProtect plug-in now includes Similarity Search on Wisenet 9 cameras: start from one image of a person or vehicle and retrieve visually similar subjects across recorded video. It sits next to existing attribute search (person, vehicle, plate, barcode) and Road AI (make/model/color/country). Live PTZ, thermal, IP audio, and IVA overlays stay in the same XProtect UI. Available now. VP Thomas Bille: open platform lets Hanwha ship camera AI at its own pace. Hanwha Europe John Lutz Boorman: operators respond from a single interface.
Why it matters
Ship Wisenet 9 + XProtect Similarity Search as the investigation add-on, not a second VMS.
Source
Milestone press release - 7
Key point
Long-form video agents that append observations drown the evidence they already found (“semantic thrashing”). VideoLoop adds an inner loop that retrieves from an unbounded filesystem of past observations and rewrites a bounded working memory after each step. Plug-and-play on four LVLM backbones: +4.2 pp average on VideoMME (long). On the hardest quartile, a blind judge reading only the agent’s context scores 81.1% vs 60.9% append-only. With Gemini 3.1 Pro: 88.3% VideoMME (long), 88.8% VideoMMMU, 80.9% LongVideoBench (long). Code: philipxjm/videoloop.
Why it matters
Multi-hour incident Q&A will ask for a rewrite memory, not a growing transcript. Design the VMS agent with a bounded context loop.
- 8
Key point
Video-RSI lets a video-understanding agent use its own LM to revise the executable harness that decides what it looks at. It re-investigates training videos to test competing failure explanations, then keeps revisions that raise accuracy and cut visual cost. Evolved agents improve accuracy while processing fewer frames. Code: bingjunluo/Video-RSI.
Why it matters
Next-year “visual AI agent” RFPs will ask who owns the observation policy. Keep the cameras; swap the harness.