IVS Briefs
IVS / Edge AI Research Brief
Overall assessment
Signal level
High on an overnight Flock expansion (East Peoria), a FY27/28 budget-kill in a Flock showcase city (Scottsdale), a 7–0 funds-and-property ban (Fayetteville), and a Flock→Axon 10-year swap (Caldwell).
Notable trends
- Overnight ALPR is expand / swap / ban / budget-kill in the same 24 hours — East Peoria grew the grid while Scottsdale, Fayetteville, Caldwell, and South Windsor wrote exits.
- Vendor swap (Caldwell Axon) is a live bid class, not only teardown.
- VMS finally has a dated SKU (Verkada Oct
- after two VMS-gap days.
Gaps
No new Hailo / Qualcomm / Jetson module SKU (8.4.174 is a Hailo INT8 calibration fix, not a board). No 48h NVIDIA Metropolis / World Labs post. Tewksbury 21–22, Fountain Hills, and Apache Junction votes are unpublished — not padded. Oklahoma Kyle order left for a Monday week catchup.
Mix counts
LPR 3 / VMS 1 / Edge 1 / Policy 2 / Lab 3. LPR+Policy city/contract = 5/10 = 50%.
Now vs On deck
Now 8 (items 1–8) · On deck 2 (CCDF, evidence-readiness).
Signals (10)
- 1
East Peoria, IL: unanimous +10 Flock cameras, $156,500 / 5 years, next door to the Creve Coeur misuse case
Key point
East Peoria City Council unanimously approved a five-year contract adding 10 Flock ALPRs on top of 18 already in the field, $156,500 over five years. About 10 protesters gathered in Levee Park; public comment ran ~30 minutes, all against. Mayor John Kahl said the city has used Flock for 4.5 years with no local misuse and that a second-reading consent item means the board had already decided. Chief David Catton said the city moved from quarterly to 30-day search audits and that authorized users must cite a call-for-service number. He would not confirm or deny whether Creve Coeur officers can query East Peoria cameras while former Creve Coeur chief Justin Egan is under Illinois State Police review.
Why it matters
This is the overnight expansion Tuesday’s brief was watching. Quote CFS-number search + 30-day audit as the local floor, and do not treat a neighbor’s misuse investigation as an automatic non-renew — this council expanded anyway.
Source
WCBU; 25News / WEEK - 2
Key point
After Chief Joseph LeDuc’s presentation, council directed staff: do not fund the Flock contract in the proposed FY 27/28 budget; identify non-appropriation / notice steps to end the contract by June 30, 2027; prepare an ALPR ordinance with criminal penalties for misuse for the Dec. 1, 2026 meeting; have the city auditor run an initial independent review, then annual audits; and evaluate alternative ALPR vendors with a transition plan that keeps public-safety coverage during a swap. SPD slide deck: 85 fixed cameras, 185-day plate retain, $280,500 annual cost, Flock program since Feb 2023, current contract through June 20, 2028. ABC15 and azfamily still had pre-meeting copy as of Wednesday morning; the YouTube closed captions are the public record of the direction.
Why it matters
A crown-jewel Flock city just wrote a budget-kill date, not an overnight pull. Bid the June 30, 2027 non-appropriation window and a Dec. 1 ordinance, and show data-ownership / no-federal-default terms if you want to be the alternative vendor they were told to evaluate.
- 3
Key point
City Council voted 7–0 (Monique Jones absent) on ordinance 2026-3024 to prohibit using city money to acquire, lease, install, operate, maintain, repair, replace, or license an ALPR system, or to buy related subscriptions, storage, derived data, or access. No ALPR may be installed or operated on city property (streets, ROW, vehicles, poles). Parking Management keeps its mobile plate readers if data is used only for lot enforcement and deleted within 48 hours unless needed for a citation (CIO Keith Macedo said the current practice is 24-hour delete). Police Chief Mike Reynolds said FPD has no desire to buy ALPRs. Sponsor Bob Stafford; Washington County passed a similar ban Sept. 17.
Why it matters
This is a preemptive funds-and-property ban, not a camera teardown — Fayetteville PD never fielded Flock. Quote the parking 48-hour / no-share carve-out if you still need a municipal plate product, and do not pitch city-funded ALPR access in a town that just wrote it out of the code.
- 4
Caldwell, ID 4–1: dump 45 Flock, stand up 65 Axon ALPR / live-feed cameras on a 10-year $6.7M bundle
Key point
Caldwell City Council voted 4–1 to extend Axon through Dec. 31, 2036 at $6.7 million (43% off vs two five-year terms at $11.29M). CPD will not renew Flock; 45 Flock ALPRs come down once 65 Axon ALPR / live-feed cameras are in, plus 4 patrol drones and 2 indoor drone kits. City Q&A: no federal agencies get default access; other locals only agency-by-agency with the chief’s sign-off; ALPR data 30 days unless saved as evidence; Axon does not get the data for product development; monthly RTIC audits. Mayor Eric Phillips had publicly backed a statewide ALPR ban, then the council still expanded the pole count under a different vendor.
Why it matters
Overnight ALPR news is a vendor swap plus expansion, not a ban. If you can bundle LPR with body-worn, Taser, and drones under one evidence store and a no-federal-default clause, that is the bid Caldwell just bought for a decade.
- 5
Key point
After a public hearing where nobody spoke in favor, South Windsor Town Council voted 6–3 to terminate the Flock contract and remove the ALPRs within 60 days. Mayor Craig Zimmerman said data will be purged and any future proposal needs new rules. A straw poll after the hearing ran 20–2 to remove (one keep vote from a police department member). Neighboring Windsor already killed Flock in July.
Why it matters
Metro Hartford is now a cover-and-return cluster with a written 60-day deprovision clock. Bid the teardown date and a local-only retain policy; do not assume a police “it works” briefing beats an empty pro-camera gallery.
- 6
Key point
Verkada’s October Command drop hits customer orgs tomorrow. Unified Timeline adds Live Mode (new sightings of a person of interest populate in near real time from on-camera processing), spotlight in crowded frames, access-control events on the same timeline, and a maps view (private beta). Same release: Catalyst MCP (Claude / ChatGPT / Gemini, private beta), Transportation SKU, CTB63 dual-spectrum thermal bullet, CR63-TE LTE telephoto remote camera, LPR+analytics on third-gen bullets, and new grab-and-run / wrong-way / obstructed-door alerts. Access-user face match in Unified Timeline follows in early November.
Why it matters
This is the VMS SKU on the truck this quarter, not another Flock council. If you already have Verkada in the stack, Live Mode + door events is the demo; if you do not, the thermal / LTE SKUs and Catalyst MCP are the competitive talk track against Genetec / Milestone this week.
- 7
Key point
Maintenance release after Monday’s 8.4.173 Xilinx Versal export. Hailo INT8 optimization now materializes a reusable calibration set before the optimizer runs, so calibration methods that revisit images no longer fail. Dynamic CoreML exports chunk batches larger than the exported size. AutoBatch backward profiling finds tensors inside nested YOLO26 heads. Fast-NMS keeps detections at IoU 0; model.track no longer leaks unconfirmed tracks. No new architecture or checkpoint.
Why it matters
If you export YOLO to Hailo boxes, this is the INT8 calibration bug that was burning compile time. Accelerated compute beats locked AI cameras: swap models vs a 36–84 month hardware refresh.
- 8
Key point
TwelveLabs released Pegasus 1.6, its first model trained natively on egocentric (first-person / wearable / teleop) video. Video-to-text returns timestamped JSON: action segmentation and labeling, dense captions, quality scoring, search/curation, and consent/compliance flags. Specs: up to 120 minutes in one call, 261,120-token context, 12 languages, native image analysis (1–20 images), zero pre-index. Aimed at robotics and physical-AI teams that currently spend 70–155 hours labeling each hour of first-person footage.
Why it matters
Wearable / robot-eye video is the next evidence class after fixed cameras. This is an API you can trial on body-worn or teleop clips this quarter, not a 36-month paper.
- 9
Key point
Baptiste Chopin, Thomas Swearingen, Arun Ross, Antitza Dantcheva, Christian Rathgeb: CCtv DeepFakes (CCDF) — 1,840 videos (460 real, 1,380 generated) across 16 crime/accident categories, generated with Grok Imagine, Google VEO 3.1, and OpenAI Sora 2. Three releases: raw, metadata-cleaned (so detectors cannot cheat on container tags), and a low-effort post-process attack set. Ten recent SOTA detectors that score well on older web-deepfake sets do not reliably separate CCDF fakes from real CCTV. Dataset is small, manually annotated, evaluation-focused.
Why it matters
Customers will start asking whether a clip from a pole camera is real. Today’s detectors are tuned on YouTube faces, not commercial video generators shooting fake crime scenes — this is the test set, not a shipping filter.
Source
arXiv:2610.07939 - 10
Key point
Dan Ben-Ami, Kobi Cohen, Chaim Baskin: streaming video-language models must decide when enough evidence has arrived, not only what to answer. A linear probe on frozen VideoLLMs reads an evidence-readiness signal (AUROC 0.733–0.905 across seven models) labelled from timestamped evidence, not from the model’s own output. The readout is question-conditioned (flips on 66.1% of byte-identical windows when only the question changes) and still works when the model answers wrong (AUROC 0.722). Readiness Gating improves accuracy by up to +9.75 pp at matched video duration with negligible compute. Trained “I’ve seen enough” triggers in released streaming systems are approximately orthogonal to this signal.
Why it matters
Live VMS agents that answer too early hallucinate; ones that wait too long miss the incident. This is the gating recipe for streaming search — research, not a Command SKU.
Source
arXiv:2610.08560