IVS Briefs
IVS / Edge AI Research Brief
Signals (9)
- 1
Key point
Cleveland Metropolitan School District will not extend Flock Safety and has directed the vendor to suspend all 201 cameras on district property. The Board of Education had been expected to vote today on a one-year $603,000 extension; staff preempted that vote after a “broad review” of privacy, law-enforcement input, cost-benefit, and existing capabilities. The original contract lapsed at the end of June 2026; cameras stayed on under a short-term extension. CMSD will keep its ~5,000 district-owned cameras, Safe Passage, CDP/RTA partnership, and temporary parking-lot monitors. A Sept. 30 audit of the Flock network: in June 2026 the system logged 141,032 searches (16 CMSD, 10,832 Cleveland PD, 59,219 other Ohio, 70,965 out-of-state); non-CDP/CMSD access was closed in July 2026.
Why it matters
This is a school-lot take-down with cameras ordered suspended, not a city ROW story. Bid parking-lot video and lighting against the 5,000-cam district stack; do not quote a live $603k Flock year.
Source
Ideastream; Cleveland19 - 2
Key point
CPD posted a draft ALPR directive (not yet in force). Videos and stills erase after 30 days unless they hold evidentiary value. External street/expressway ALPRs are not city-owned (many are Flock); query access to those would require bureau-chief approval, with each search traceable to the submitting member and random audits. The draft restates Illinois’ 2024 ban on sharing for reproductive-health or immigration enforcement. Fox 32: CPD officers submitted more than 23,000 Flock queries May–July 2026 under the city’s MOU. Public comment is open until Nov. 12. Illinois State Police run more than 800 ALPRs.
Why it matters
Chicago is writing query-governance on a third-party Flock mesh, not buying city cameras. Any CPD/ALPR bid now has to survive 30-day retain, bureau-chief search, and a Nov. 12 comment clock.
Source
Chicago Tribune - 3
Key point
City Attorney Gregory Venker’s seven-page Oct. 24 opinion: council cannot dictate how LPD uses ALPR evidence, but it can prohibit ALPR placement in city rights-of-way. 1st Ward Ryan Kost asked for ordinance language on that ROW chapter. LPD leases 26 Flock cameras at seven locations. An Aug. 24 resolution asked Mayor Andy Schor to stop using ALPRs; the administration did not act. Council President Peter Spadafore (the only no on that resolution) now wants a full ordinance so parking and school-bus readers are scoped. The city says the Flock lease has about a year and a half left; Kost said he would “unplug” them anyway. Flock told council in July that about 140 other in-state departments could see Lansing cameras.
Why it matters
The live path is a ROW placement ban, not a PD evidence rule. Quote the Venker opinion when a Michigan city says council “cannot touch” leased Flock poles.
Source
WKAR - 4
Key point
Oklahoma’s ALPR statute already limits retained plate data to Compulsory Insurance Law enforcement — then adds “or as otherwise permitted by law.” Sen. Mary Boren (D-Norman) brought a Senate Public Safety Committee interim study Oct. 21 to tighten those six words before the winter session. Presenters walked the panel through GPS/cell-site warrant cases and district-court rulings that agencies exceeded insurance-only use (including a McClain County 2024 evidence bar). Edmond and The Village chiefs attended; neither spoke. Boren: “People should feel secured by their government, not stalked by it.”
Why it matters
Oklahoma ALPRs are already insurance-purpose on paper. A winter rewrite of the six-word clause is the RFP risk for any Flock or Motorola stack that treats insurance as a gateway to general investigative search.
Source
The Oklahoman - 5
Key point
MYIR’s MYC-JR3576 is the socketed SMARC 2.2 (82×50 mm, 314-pin MXM) counterpart to the 2024 solder-down MYC-LR3576. SoC: RK3576 / industrial RK3576J — four Cortex-A72 @ 2.2 GHz + four Cortex-A53 @ 1.8 GHz, Mali-G52 MC3, 6 TOPS INT8 NPU (INT4/8/16, BF16, TF32). Decode 8K30 or 4K120; encode 4K60 H.264/H.265. I/O: 3× MIPI CSI, 2× GbE, USB 3.0, PCIe 2.1 or SATA, optional Wi-Fi 6E/BT 5.4. Memory 4 or 8 GB LPDDR5, 32 or 64 GB eMMC. Linux 6.1 (Debian/Yocto) and Android 15 (customer BSP). Sample pricing: commercial 8 GB/64 GB $209; industrial 4 GB/32 GB $192; MYD-JR3576 kit $279 / $293. Modules and kits shipping now.
Why it matters
This is a socketed 6 TOPS camera/NVR brick you can buy this week, with three CSI ports and 8K decode, without a locked camera SoC. Accelerated compute beats locked AI cameras (swap models vs a 36–84 month refresh).
Source
CNX Software - 6
Key point
Shuyang Yu et al. present OVM-Occ: Language-Embedded Gaussians, geometry-aware deformable attention, instance-level vision-language embeddings, and decoupled semantic distillation for open-vocabulary indoor occupancy from a single RGB frame. New eval sets OpenOcc-ScanNet and OpenOcc-Replica. They report +12.2 mIoU and +0.6 IoU vs prior SOTA on Occ-ScanNet, with consistent gains under open-vocabulary protocols.
Why it matters
Occupancy that is not locked to a closed class list is the indoor perimeter / lobby path when a VMS RFP asks for “what is in this volume” without a site taxonomy.
Source
arXiv:2610.35312 - 7
Key point
Mayur Kulkarni’s RoR splits traffic-anomaly VLMs: a LoRA-tuned 9B specialist on structured, temporally grounded chain-of-thought, a 4-bit 27B specialist on answer style, and a training-free LLM judge that picks per question. 1st on the in-domain Traffic Anomaly Reasoning leaderboard, mean score 0.6788. Closed-form questions reward evidence; open-ended answers reward annotation style — the two specialists stop one model from doing both poorly.
Why it matters
This is the collision / violation / hazard write-up path cities now put in video-analytics RFPs: detect, describe, and explain, not a bounding box.
Source
arXiv:2610.34993 - 8
Key point
Yihao Hu et al. argue action-recognition labels under-specify event order, count, and relations. Visible-State Supervision (ViSS) trains on source-grounded probes of identity, order, accumulation, relation change, and answer sufficiency. Test-Time Causal View Search (TVS) then expands from recent frames to sparse event/history views on demand, keeping the model fixed. Together they report the best published accuracy among compared systems on StreamingBench and OVO-Bench at a fraction of exhaustive-frame compute; the largest ViSS gain is hallucination detection. Strict replay shows most of the remaining oracle gap is a correct answer later overwritten.
Why it matters
Live-stream VLMs that forget a correct alert are the failure mode on an NVR. Quote ViSS when a customer asks why the model “saw it and then unsaw it.”
Source
arXiv:2610.34926 - 9
Key point
Shih-Yao Lin’s EARS is a training-free, query-free front-end: rank frames by optical-flow motion and audio energy, then temporal NMS for diversity. Evidence score alone is frame AP 0.39; used only to choose VLM frames it reaches localization AP 0.94 vs 0.88 uniform. At K=1, XD-subset F1 0.660 vs 0.614 centered uniform. On mostly-silent UCF-Crime it slightly trails uniform. Code: github.com/mikeslin-ir/EARS.
Why it matters
This is how you put a VLM on a gunshot / crash stream without sampling 30 fps. Use it when the event makes sound or motion; stay uniform on stealthy UCF-style footage.
Source
arXiv:2610.35536