IVS Briefs
IVS / Edge AI Research Brief
Signals (10)
- 1
Key point
Inola Town Council voted unanimously to end its Flock ALPR contract. Renewing would have cost $21,000 (NewsOn6) / $21,600 (Radio Oklahoma News). State Rep. Tom Gann (R-Inola) spoke against a statewide/national “mass surveillance dragnet.” Resident Steven Flathers wants the line reallocated to police. Leaders instructed the town attorney to contact the company on next steps. No removal date is announced. Radio Oklahoma: the vote does not mean the contract is already terminated, and recovery of a renewal payment is uncertain. Coweta’s August non-renew is separate context.
Why it matters
A unanimous small-town unwind with hardware still live is the Oklahoma pattern after Coweta. Bid a local VMS against a ~$21k Flock year; quote attorney-contact / no-date-remove language until a PO actually dies.
- 2
Key point
Simpsonville City Council voted 5–2 to place Ward 4 Councilwoman Sherry Roche’s ALPR-ban ordinance on a future agenda for first reading. The draft would ban ALPRs in city limits and terminate associated contracts. First reading is expected at the November meeting; two more readings are required before it becomes law. Yes: Roche, Chad O’Rear, Aaron Rupe, Shannon Williams, Jerry Tuso. No: Mayor Paul Shewmaker and Ward 5 Tom Pinkerton. Chief Mike Hanshaw: the city runs 12 cameras on Flock’s original plate-only model (not newer AI units); he cited a fatal hit-and-run ID. This is an advance to first reading, with cameras still up.
Why it matters
Upstate SC now has Simpsonville on a three-reading calendar beside Anderson County’s already-advanced ordinance. Price a 12-camera local retain/VMS bid against a November first reading; do not treat the 5–2 as a teardown.
- 3
Key point
Akron opted out of nationwide searches of its Flock network, so outside agencies can no longer hit Akron cameras via blanket national queries. Mayor Shammas Malik: those searches bring “little to no public safety benefit” and open uses that conflict with city policy. Public Safety Chief Craig Morgan: the city will keep reviewing logs. Ohio agencies can still run statewide searches. Outside agencies with a suspect believed to be in Akron must request Akron PD directly. Immigration and reproductive-health search categories for outside agencies were already disabled in May. Internal-use review continues. Beacon Journal (paywalled) later reported the Citizens’ Police Oversight Board’s Oct 28 special meeting: 7–1 against a suspend-until-audit letter, then a no-renew pending independent audit letter (contract through Dec 31 2028) that the governance committee will still edit — letter not sent. Do not treat that as a city terminate.
Why it matters
A nationwide-off / statewide-on toggle is the live Akron control, with a 2028 contract still in force. Ohio ALPR bids now have to price state-network access separately from national Flock search.
Source
Spectrum News 1 - 4
Edmond unanimous Avigilon appearance-search cameras at Arcadia Lake; ~$90k + $30k license; FR license not bought
Key point
Edmond City Council unanimously moved $250,000 from park reserves into the Arcadia Lake master plan after the May 3 shooting that killed Avianna Smith-Gray and injured 22. Interim Parks Director Frank Hill: about $90,000 covers cameras and on-site server; $30,000 recurring licensing every 3–4 years; $130,000 is automatic gates and other upgrades. In September the city awarded $89,327.87 to Oklahoma City Digi Security Systems for Avigilon (same stack as City Hall, Mayor Mark Nash). Operators can search by appearance (race, gender, age, build, clothing) and click a box to trace a person across cameras. Avigilon face recognition needs a separate license the city has not purchased. Staff: cameras face outward at five lake entrances (not the interior); install target six months. These are CCTV, not ALPRs.
Why it matters
Parks RFPs after a mass-casualty event are awarding Avigilon Unity appearance search with the FR SKU left off. Quote Digi’s $89k camera/server line and a 3–4 year license against Flock-style vehicle search; keep face recognition as an optional license, not a default.
Source
The Journal Record - 5
Milestone XProtect 2026 R2 + Arcules fall update: maps, sticky notes, server logs, edge-gateway cloud backup
Key point
Milestone shipped XProtect 2026 R2 with autumn Arcules updates. CTO Andrew Burnett: operators need context; admins need a way to run growing hybrid estates. Maps puts cameras, alarms, devices, and live video on geo/floor-plan views (campuses, hospitals, airports); Maps lands in a future Smart Client later in Q4. Sticky Notes attach shift/handover text to a camera view, managed in Management Client. Server Logs centralizes audit/system/rule logs and can feed third-party SIEM. Permission Inheritance applies group ACLs to newly added cameras. XProtect Remote Manager (Care Plus) is browser multi-site health. Arcules: Cloud Backup for Edge Gateways (lower-res secondary stream), Triton Sensors occupancy/environment into alarms, automatic HTTPS device communication, faster local streaming, hybrid upgrade/diagnostics and role-based alarm access.
Why it matters
This is the first VMS SKU drop in weeks. Campus/hospital RFP language can now name R2 maps (Q4 Smart Client), centralized logs, and edge-gateway cloud backup instead of a camera-locked cloud VMS.
Source
SourceSecurity - 6
Key point
LattePanda’s Sigma 2 is a 146 × 102 mm 3.5-inch SBC on Intel Core Ultra X7 358H (Panther Lake-H): 16 cores (4P+8E+4LPE), Arc B390 (12 Xe, 122 TOPS INT8), 50 TOPS NPU, combined 180 TOPS, 64 GB LPDDR5X-8533, 1 TB NVMe, dual 10GbE (RTL8127), 3× M.2 M-Key (PCIe 5.0 x2 / 4.0 x4 / 4.0 x1), OCuLink, two Thunderbolt 4, RP2350 GPIO MCU. Ubuntu image ships OpenVINO Model Server with Qwen3.6-35B-A3B INT4 (~30–34 tok/s); up to 59 GB can be allocated as VRAM. Pre-order $2,599 through mid-November, then $2,799. Lifecycle ≥3 years. 0–40 °C; idle ~4.5 W; ~38 W in a Qwen3.5-4B GenAI test.
Why it matters
A $2.6k / 180 TOPS / 64 GB box is an on-prem VLM/analytics node you can swap models on, with dual 10GbE for multi-camera ingest. Accelerated compute beats locked AI cameras (model swap vs 36–84 mo refresh).
Source
CNX Software - 7
Key point
Xiangyu Liu et al. train RGB + thermal/event/depth trackers without pooling raw video. The pretrained backbone and head stay frozen; clients exchange prompt parameters only. The server keeps a global RGB bank and separate thermal, event, and depth banks. RGB-only clients still update the RGB bank and are kept out of fusion/normalization aggregates. Mean vs FedAvg: +2.35 AUC LasHeR, +3.45 SR VisEvent, +2.11 F-score DepthTrack (three matched repetitions, fixed partition). Highest mean thermal AUC and depth F-score among nine methods.
Why it matters
Multi-sensor perimeter (RGB + thermal + event) can federate prompts instead of shipping video to a central trainer. Quote frozen-backbone RGB-X when a city or campus will not pool footage.
Source
arXiv:2610.34939 - 8
Key point
Muxi Zha et al. (IEEE T-IP) keep event data in a 2D format so image MOT still applies, then add extra spatio-temporal views to split overlapping trajectories. MvP encodes three complementary views; MvF lifts tracklets into 3D and associates them with an optimization. Evaluated on four datasets including a self-collected Small Objects Dataset. Training-free, so it runs where event-stream labels do not exist. Code: github.com/zhazhabiu/MvTracking.
Why it matters
Event cameras are the high-speed / high-dynamic-range perimeter path. A training-free multi-view event MOT is the association layer to put on an event + RGB RFP when you cannot label the site.
Source
arXiv:2610.37309 - 9
Key point
Jingong Chen et al. drop the pinhole assumption inside geometry foundation models. Positional encoding and the training loss are conditioned on the dense ray field (unit direction per pixel), so fisheye, wide-FOV, and mixed lens types can share one feed-forward reconstructor. Angular rotary encoding plus solid-angle loss weighting are lens-agnostic. On a 13-camera / 3-dataset benchmark using each method’s own predicted poses, Any3R is the most consistently accurate. Project: jingong-chen.github.io/Any3R/.
Why it matters
Site meshes mix dome, bullet, and fisheye. A single feed-forward reconstructor that does not retrain per lens is the digital-twin / occupancy path for heterogeneous camera farms.
Source
arXiv:2610.36150 - 10
Key point
Xusheng Liang et al. argue long-video QA fails less from context size than from shortcut bias (linguistic correlations + salient-but-irrelevant frames) under sparse sampling. They introduce S-MME to diagnose that bias, then CLEAR: treat the full video-question pair as the complete view, estimate bias from question-only and locally salient clips, and intervene hidden states at inference so the right answer can win. Training-free; plugs into multiple MLLMs; gains on five benchmarks. Code: github.com/Xiaoqiovo/CLEAR.
Why it matters
Control-room VLMs sample hours of video at a few frames. A test-time bias intervention is the evidence-readiness layer to quote when an RFP asks a VLM to find the decisive moment without a fine-tune.
Source
arXiv:2610.36210