IVS Briefs
IVS / Edge AI Research Brief
Signals (8)
- 1
Alexandria 61 Flock ALPRs deactivate today; hardware out by December; Nov 10 is a rubric, not a rebid
Key point
Alexandria Police Chief Tarrick McGuire: “I anticipate by this Friday that the ALPR camera should be deactivated.” “The goal is to have them all removed by December.” City Manager Jim Parajon: the move follows council direction to end the Flock contract and stop using the company’s 61 cameras. Mayor Alyia Gaskins: “Flock as a vendor needed to be out.” Vice Mayor Sarah Bagley: Nov 10 is a rubric for evaluating products/vendors, not a day for Flock to propose a new contract. Red-light and school-zone speed cameras stay unless directed otherwise. APD started with Flock in 2023. ALXnow: nearly 21,000 queries Aug 2025–Aug 2026 → 72 felony + 65 misdemeanor charges, 3 missing persons, 35 stolen vehicles, 4 firearms. Do not claim hardware already removed.
Why it matters
A Northern Virginia capital-city go-dark today / poles still up unwind. Bid a local VMS against 61 city-owned readers on leased poles, with a December physical-remove window and a Nov 10 surveillance-eval rubric — not a Flock rebid.
Source
WTOP; ALXnow - 2
Wausau 9–1 orders 20 Flock cameras covered in two weeks / Veterans Day; ~$2,500 cover; $50k 2027 line already cut
Key point
Wausau Common Council voted 9–1 to cover all 20 Flock ALPRs and stop police use of the data. Flock owns the cameras, so the city cannot alter them. Mayor Doug Diny: staff laid out a two-week window; every camera should be covered by Veterans Day; resolution language is written so the police chief does not have to climb a ladder. Cover cost: employee wages + equipment ~$2,500. Council already stripped the $50,000 2027 Flock line. Pilot: Alder Sarah Watson wanted them covered “tonight”; Public Health and Safety had recommended 5–0 on Oct 19; the item was added late and needed a two-thirds rule suspend. Five-year $250,000 contract (Oct 2023) still sits under the cover order. Legal question whether covering is allowed remains live.
Why it matters
Central Wisconsin now has a cover-not-tear-down 9–1 with a Veterans Day hardware deadline and a $50k budget cut already on the 2027 book. Quote a 20-camera local retain/VMS against ~$2,500 bags plus the remaining contract years; do not treat the vote as cameras already dark.
- 3
Key point
Mayor Yvonne Flowers “made good on her promise to veto” the Common Council’s vote to end all Flock contracts (LPRs and gunshot detection). Only Councilman Nathan Shook opposed the original resolution. Veto message: this is not a pro-Flock stance; the legislation has public-safety, operations, contract, and budget implications that need a collaborative review. Flowers says she was blindsided when Councilmembers Daniel Atonna and Evan Menist announced the plan at the Dutchess County Legislature rather than at city council. Planned 2027 speed and red-light cameras are linked to the existing LPRs and gunshot detection. North-side resident Flowers cites crime drop and convictions from Flock evidence. Evan Menist did not respond before publication. Distinct from the 10-23 numbered 8–1. Override watch Nov 2.
Why it matters
The Hudson Valley now has a live veto, not just an 8–1 cancel. Price a local VMS / gunshot alternative against a Nov 2 override; keep 2027 speed/red-light cameras on the same stack conversation.
Source
Mid Hudson News - 4
Solink raises >$100M (Vertu-led) for AI vision intelligence on existing cameras; 500k cams / 35k sites
Key point
Ottawa cloud-video / AI vision company Solink raised more than $100 million, led by Toronto Vertu Capital, with material participation from Growth Equity at Goldman Sachs Alternatives and new money from Golub Growth and Comcast Ventures. ~90% primary / rest secondary (Globe and Mail / BetaKit). Total funding now >$198 million. Platform: 500,000 cameras at 35,000+ locations in 65 countries, on cameras customers already have — no proprietary cams. CEO Mike Matta: a “watershed moment” for physical businesses. Revenue growing >40%/yr; Matta told the Globe the company is on track to surpass $100 million ARR by April. Feb AI agents for perimeter/loss prevention on edge devices; March VerifEye VLM false-alarm filter. Customers include Take 5, Jack in the Box, PGA TOUR Superstore. Goldman led the $60 million Series C in July 2023. Valuation undisclosed.
Why it matters
This is the first large cloud VMS / vision-intelligence round in weeks, aimed at retail/restaurant/warehouse estates that keep their existing cameras. Quote Solink’s edge-agent + VerifEye VLM stack against a camera-locked SOC when an MSP is bidding multi-site loss-prevention without a forklift camera swap.
Source
Solink; SiliconANGLE - 5
Key point
Allen Tu et al. present GOFish, the first 3D Gaussian Splatting surface-reconstruction method that trains directly on fisheye images, rasterizes FoVs beyond 180°, and extracts an aligned mesh from the same scene. Prior fisheye 3DGS only rasterized to 180° and produced no mesh; Gaussian surface methods assumed pinhole training images. Calibrated densification, Calibrated Mip Filtering, and Wide-Angle Appearance Modeling keep pixel blur and lens falloff out of the scene and mesh. +2.28 dB PSNR on 200° FIORD and +1.97 dB on 205° FullCircle vs the best prior; highest mesh F-score on FIORD and ScanNet++ main. The Gaussian scene + mesh is a LiDAR-free, simulation-ready digital twin with real-time render, full-FoV ray tracing, and collision geometry. Project: 3dgofish.github.io.
Why it matters
360° / fisheye perimeter cameras are already on poles. A LiDAR-free mesh from those lenses is the occupancy / digital-twin layer to quote when a campus RFP asks for collision geometry without a second sensor truck.
Source
arXiv:2610.37818 - 6
Key point
Lu Wang et al. argue proactive streaming video agents that can only wait or answer will wait forever or hallucinate when video or user information is incomplete. They add a third action — ask — and introduce AskStream, the first multi-turn benchmark spanning video and user uncertainty: 4 categories / 8 subtasks, 438 videos, 2,714 frame-level wait/ask/answer windows. Across multiple-choice and open-ended settings, even the strongest model reaches only 52.18% Ask-F1. Project: askstream.top; code github.com/wanglu-cs/AskStream.
Why it matters
SOC copilots that auto-answer on a live stream will invent the missing plate, face, or camera. An ask-before-answer gate is the RFP language to put on streaming VLMs so operators get a clarifying prompt instead of a hallucinated incident.
Source
arXiv:2610.38678 - 7
Key point
Christoph B. Rist, Larissa T. Triess, Markus Enzweiler, and Dariu M. Gavrila (IEEE Intelligent Vehicles Symposium 2026) train LiDAR semantic segmentation with an occupancy-prediction auxiliary on aggregated source scans, then drop that head at deploy. Input stays a single-frame LiDAR scan at train and test — no target scans, no target labels, no sensor priors, no extra inference cost. Supervision via a simple world model of accumulated source points, replacing single-scan segmentation training. On unseen sensors, generalization improves in some setups with competitive source-domain scores.
Why it matters
Perimeter / occupancy stacks mix Ouster, Hesai, and Velodyne units over a 36–84 month refresh. A source-only occupancy auxiliary is the transfer story to quote when a site adds a new LiDAR SKU without a labeled campaign.
Source
arXiv:2610.39516 - 8
LiDAR-SHIELD: Dockerized adversarial robustness tests for 3D perception (HTTP
/info/forward/gradient)Key point
Alexandros Gkillas et al. ship LiDAR-SHIELD, a fully Dockerized framework that stress-tests LiDAR 3D perception without installing the target model inside the toolkit. Any model is exposed through three HTTP endpoints — /info, /forward, /gradient — so a new backbone or dataset needs only a loader/adapter. Same engine covers 3D semantic segmentation (PGD, Square Attack) and 3D object detection (point perturbation, removal, injection, LiDAttack). Code: github.com/alexandrosgk/LiDAR-shield.
Why it matters
Occupancy and perimeter LiDAR bids now have a vendor-neutral attack harness. Ask for a LiDAR-SHIELD (or equivalent) report before locking a 3D detector that will sit on a fence line for years.
Source
arXiv:2610.39848