IVS Briefs
IVS / Edge AI Research Brief
Overall assessment
Signal level
High on a first-in-OC six-month Flock pause, a 7–0 Arizona ordinance that criminalizes surveillance misuse, and a Massachusetts town meeting that referred ALPR bylaws instead of passing them.
Notable trends
- Overnight ALPR is pause / criminalize-misuse / refer-the-bylaw, not another bag-and-tag. Fullerton keeps the 2028 contract and runs a six-month A/B; AJ writes Class 1 + job loss; Tewksbury already killed Flock (10-
- and parked the next-vendor rules at 2027 ATM.
- Edge this window is a Thor VPX SBC, not a Hailo/DeepX module.
- Lab is surveillance VLM + VAD/VAA + high-dynamic streaming + onboard detect-then-crop.
Gaps
No 48h VMS SKU (Verkada Live Mode already numbered 10-07; Genetec Carmel City is a weak case-study PR). No new Hailo / Qualcomm / DeepX module SKU (WOLF-16T9 fills Edge on IGX Thor). Slot3R streaming 3D (arXiv:2610.12282) left unnumbered to hold mix. Milford CT Tech Review Thu 10-08 had no vote (recs in 60–90 days; next Nov. 10) — Condor pause already honorable 10-07. West Lafayette “unanimous Resolution 25-2026” recircs are the 10-05 first vote numbered 10-06; final is Nov. 2. Fountain Hills Oct 6 was 5G/wireless, not Flock. Ultralytics still 8.4.174 (numbered 10-07).
Mix counts
LPR 1 / VMS 0 / Edge 1 / Policy 2 / Lab 4. LPR+Policy city/contract = 3/8 = 38%.
Now vs On deck
Now 6 (items 1–6) · On deck 2 (FastBench, Versal maritime).
Signals (8)
- 1
Key point
Fullerton City Council voted 3–1 to suspend 19 Flock ALPRs (17 fixed + 2 mobile) for six months. Yes: Ahmad Zahra (motion), Shana Charles, Nicholas Dunlap. No: Jamie Valencia. Mayor Fred Jung absent. Dunlap’s amendment lets staff resume immediately for contract disputes or severe public-safety emergencies. The pause is a with-vs-without test of investigation-completion rates, plus a town hall and a review of Flock audit protocols and the city’s ALPR policy. Contract runs to 2028; Voice of OC: $51k/yr to operate, about $236k spent since the 2021 Flock deal. Chief Jon Radus. Santa Ana kicked its ALPR discussion to Oct. 20.
Why it matters
Orange County now has a timed A/B pause, not a teardown. Bid the six-month window with local-only retain, named-supervisor search, and a 2028-contract rewrite; Santa Ana Oct. 20 is the next OC calendar date.
- 2
Key point
Council passed Ordinance 1579 7–0. Misuse of any police surveillance tech — ALPRs (Flock or other), drones, RTCC, and body cameras added at the Oct. 5 work session — is a Class 1 misdemeanor (up to 6 months jail + $2,500) and job loss. Chapter 17: plate data 30 days, no selling, monthly audits with a check outside PD. Access is limited to PD employees designated by the chief; each search needs login + case/incident number + named-supervisor authorization. Three allowed uses: hot-list alerts; misdemeanor/felony investigations; missing/endangered/vulnerable people. Sharing to other cities/towns/counties needs supervisor pre-approval under the same rules. Annual privacy/legal/ethics training with a signed acknowledgment. Companion Resolution 26-43. Thirteen speakers asked for a warrant and public audit results — not in the code. Chief Pooley: Flock misreads happen; human eyes on the plate before acting; an alert is not probable cause. Follows an April officer personal-search resignation. City spend ~$80k ALPR vs ~$120k fully loaded officer.
Why it matters
Pinal County just made misuse a crime, not a policy memo. Put CFS-number search, named-supervisor ACL, 30-day retain, and an outside-PD audit in the SOW; neighboring councils will copy the Class 1 language.
Source
Pinal Post - 3
Key point
About 160 voters did not accept or reject the bylaws. Petitioner Select Board member James Mackey moved to refer both articles to a new Surveillance Technology Advisory Committee “for review in preparation of updated proposed bylaw text” at the 2027 Annual Town Meeting. Voice votes approved the referral. STAC is seeking applicants by Oct. 16. Article 21 = surveillance oversight (define the tech; prohibit wireless-device IDs and biometrics; Select Board approval before acquire; data ownership). Article 22 = ALPR limits (Flock or Axon/Motorola): documented reason for a lookout list and limits on outside-agency access. Flock cameras on Shawsheen St are already bagged after TPD ended the relationship (Chief Ryan Columbus) — the terminate was numbered 10-01; this is the bylaw-referral result. Mackey: ~140 residents signed petitions in a four-hour rain-day block; ending Flock does not settle the rules for the next system. Select Board John Lombardo (who proposed STAC) prefers committee autonomy over highly restrictive bylaws.
Why it matters
The cameras are dark; the rules for the next vendor just slipped to 2027 ATM. If you bid Tewksbury, you are selling into STAC (apply-by Oct. 16), not a live Flock replacement this quarter.
- 4
Key point
WOLF Advanced Technology posted the WOLF-16T9 (VPX3U-THOR-RD), a SOSA-aligned 3U VPX rapid-development SBC around NVIDIA IGX Thor (T5000). Specs: 2,560 CUDA, 96 5th-gen Tensor, 14-core Arm Neoverse V3AE, 128 GB ECC LPDDR5X, 1 TB NVMe, PCIe Gen5 x4, dual 10GBASE-KR, 2.5GbE, MIG, 60–140 W. Video: up to 6× 4Kp60 H.265 encode / 10× 4Kp60 decode; Isaac, Metropolis, Holoscan. MES quotes 2070 FP4 TFLOPS. Vendor page: −40°C to +70°C standard (operational to +80°C); 40G shock / 10G sine; slot profile 14.2.16; lifecycle preliminary.
Why it matters
This is a swap-the-weights Thor box in a VPX slot, not a locked camera SKU — C5ISR and industrial sites can run Metropolis/Holoscan and change models without waiting on a 36–84 month camera refresh. Accelerated compute beats locked “AI cameras” (model swap vs 36–84 month refresh).
- 5
Key point
Saverio Cavasin, Pietro Tedeschi, Mattia Tamiazzo, Alessandro Brighente, Simone Milani, and Mauro Conti propose VESSI (VLM-Enhanced Support for Surveillance and Investigations): prompt-driven interrogation of video, with salient visual features converted to text. Four SOTA VLMs on unlabeled sets. They introduce a reference-free Composite Model Utility Score (CMUS). The most reliable model flagged potentially relevant activity in >66% of videos and cut review time by >85%. CMUS ordering matched the false-positive-rate ranking from 5,909 manually referenced frames and matched a normal-video CMUS run. ACM I.2.10 / I.4.8.
Why it matters
This is a review-time VLM for existing CCTV, not a new camera. Quote the CMUS ranking and the >85% cut when a customer asks whether a VLM in the VMS actually reduces analyst hours.
Source
arXiv:2610.11674 - 6
Key point
Congqi Cao, Zhenhe Liang, Hanwen Zhang, Yifan Zhao, Qinyi Lv, Lingtong Min, and Yanning Zhang unify video anomaly detection and anticipation in Uni-DSM. A shared noise-conditioned score transformer with scene-dependent embeddings and motion-aware weighting. VAD uses autoregressive DSM (ADSM) that accumulates anomalous evidence by denoising. VAA adds a lightweight auxiliary decoder and self-distilled DSM (SDSM) that supervises from output discrepancies because future ground truth is unavailable. SOTA on both tasks at high efficiency.
Why it matters
Most VAD SKUs are post-hoc. Anticipation (SDSM, no future GT) is the 36–84 month path from “flag the clip” to “warn before the event.”
Source
arXiv:2610.11149 - 7
Key point
Yuxuan Hu et al. (CUHK MMLab) introduce FastBench: 306 QA pairs across eight domains and six capabilities, with forward/instant/backward scopes and human-annotated evidence intervals, built so questions answerable at 2 FPS are filtered out. Strongest model Gemini-3.5-Flash scores 50.7%. Denser sampling lifts Qwen3-VL-8B from 32.9% at 2 FPS to 44.6% at 24 FPS, then saturates as history is compressed. Training-free ProactiveFrame (dual-tier sliding window: recent high-FPS + sparse history) beats sparse uniform sampling by 5.4 / 1.5 points and still sits well below oracle-guided focusing. Code: github.com/Ashone3/FastBench.
Why it matters
Live VLM-in-VMS demos that sample at 1–2 FPS will miss fights, plate flashes, and weapon draws. Ask vendors for high-FPS evidence intervals, not a 2 FPS chat overlay.
Source
arXiv:2610.12427 - 8
Key point
Thomas Goudemant, Aurélien Bobey, Omar Hlimi, and Marjorie Bellizzi train YOLOX-S on a controlled set from 68 annotated Maxar scenes (43 vessel classes), quantize it onto the DPU of a Versal VC1902, and measure a few seconds per scene with limited detection-quality loss. Downlink modes range from metadata-only (box/class/score) to crops, tiles, or a full scene with a degraded background. On dense harbor/coastal scenes, tiles keep 98% of vessels at 29% of scene volume; crops keep 83% at 3%. Complementary to conventional onboard compression. ACM I.4.8 / I.2.10 / C.3.
Why it matters
Same pattern as site NVR bandwidth: detect-then-keep-the-crop, not encode every pixel. Versal DPU is the box you swap models on. Accelerated compute beats locked “AI cameras” (model swap vs 36–84 month refresh).
Source
arXiv:2610.12182