IVS Briefs
IVS / Edge AI Research Brief
Overall assessment
Signal level
High on Lexington’s 7M / 96% outside search log, Houston’s 224/553 grant dark, and a real Europa inference card in Dell/Supermicro. PA governor is open to a ban.
Notable trends
- Abbott freeze is now two big-city partial darks (Dallas 321 delayed; Houston 224).
- The sharing graph is measurable: Houston/Dallas outrank the host city on Lexington.
- PA is ban-talk + FOIA-of-audits, not just a city vote.
- Edge news is a 45 W PCIe AIPU with H.265, not another locked dome.
- Lab is VLA cache on Thor and ReID through occlusion.
Gaps
No new Hailo / DeepX / Jetson module SKU (Thor appears only inside the rMuscle paper). No new Milestone/Axis VMS SKU (TEPCO is a Genetec deployment). Dallas budget did not produce a separate Flock roll-call this morning. Montclair ALPR-safeguard story did not yield a usable primary.
Mix counts
LPR 2 / VMS 1 / Edge 1 / Policy 2 / Lab 2 (LPR+Policy = 4/8 = 50%).
Now vs On deck
Now 6 (items 1–6) · On deck 2 (rMuscle, Tracking the Unseen).
Signals (8)
- 1
Key point
Chief Noe Diaz: Abbott’s MVCPA pause will cost HPD more than 200 cameras. Chronicle: 224 of 553 were bought with the statewide $1 auto-insurance fee. Officers run tens of thousands of searches a month. Loss hits neighborhoods without private/management-district poles. KPRC: current-year MVCPA grant not spent on Flock yet; a prior grant still runs ~200 live cams; new grant money may go to other kit. Harris County SO: $150k grant likely redirected; they run a larger Flock net bought by other county entities. Abbott freeze estimated up to 3,200 cams statewide.
Why it matters
Same freeze as Dallas, different city: Houston is partial dark, not an exit. Map grant poles vs city poles vs HOA/Home Depot. Quote Diaz’s “tough decisions” in every Gulf Coast bid.
Source
Houston Chronicle; KPRC - 2
Key point
Taxpayers fund 125 Lexington PD cameras. Audit of every search: >7 million; LPD itself ~63,610. ~96% from out-of-state or federal. Top searchers: Houston PD ~349,535, Dallas PD ~297,199, Gwinnett Co. GA ~108,592, Louisville ~78,514, Hillsborough FL SO ~69,212, Michigan SP ~68,157, Texas DPS ~65,307 — then Lexington 8th. Connecting is not automatic: outside agency requests, LPD manually approves. Flock UI lets an officer search local or nationwide. Cdr. Matthew Greathouse: bigger cities may just default to nationwide. Logs do not say whether a search returned a hit.
Why it matters
This is the sharing-graph job in one table. Dallas just cut share-out to ~200 TX — and still ranks #2 on Lexington. Put who can query us / who we query in the RFP, not a privacy slogan.
Source
WKYT - 3
Key point
Crowd map: ≥238 plate readers across Allentown, Bethlehem, Hanover Twp, North Whitehall (not all Flock — Motorola/Axon too). Allentown: 67 ALPRs + downtown gunshot detectors; $1.5M 2024 state grant. Chief Charles Roca: legitimate investigative use. Morning Call + Reporters Committee sued for Allentown search audits; city calls them confidential criminal records. Councilmember Jeremy Binder: open Flock discussion late September; wants an outside consultant and a surveillance-oversight ordinance. Bethlehem: ≥20 since ~2021; Chief Michelle Kott: case-number searches, quarterly audits, no documented misuse, no active DHS share; launching a public transparency portal (counts, not full logs). Gov. Josh Shapiro “real concerns,” open to a statewide ban; GOP Treasurer Stacy Garrity also against. Harrisburg bills: ban, warrant-to-search, misuse penalties — none on the floor. Flock’s 7-day default remains optional.
Why it matters
PA is now governor-open-to-ban + records-lawsuit + council ordinance. Bid the audit-log public-records clause. Don’t assume Flock’s 7-day retain is on unless the contract says so.
Source
Morning Call - 4
Key point
Council ordinance to continue Flock 6–1 (4th Ward Royal Mayo no). Mayo: a constituent told him the cameras had been up since September 2025 — council found out after install. Petition circulating; Mayo: a little over 500 signatures and “the cameras would have to be temporarily shut down until it’s over,” then a ballot if council doesn’t kill them. Mayor Ralph Petrella supports the cams, understands data-handling concerns.
Why it matters
Ohio keep-path can still become a citizen referendum. Ask whether hardware went live before the ordinance. Quote the petition-to-dark trigger in any small-city close.
Source
WTOV - 5
Key point
Tokyo-area TEPCO Power Grid picked Genetec Security Center as the spine for 1,600 substations/service centers and >30,000 cameras — remote monitoring, incident response, equipment management. Multi-vendor cameras kept; SDK for custom analysis. Open architecture so existing hardware stays. Critical-infra digital-transformation pitch, not a new camera SKU.
Why it matters
Utility/substation RFPs this quarter: unify 30k mixed cameras, don’t rip. If you sell Genetec, this is the reference architecture to drop in the deck.
- 6
Key point
Dutch Axelera’s second AIPU, Europa: 8 cores, 200 GB/s, claimed 629 TOPS in a 45 W envelope; RISC-V vector pre/post + on-chip H.265 so surveillance decode doesn’t bounce to the host. Cards: Edge 232p (HHHL) in validated Dell XE5 and Supermicro 111AD; Server 250p FHFL dual-slot, 128/256 GB LPDDR5, up to 8 cards/server. Voyager SDK spans Metis (edge) and Europa (server); Wingman NL pipeline builder. CEO Fabrizio Del Maffeo: >600 customers, deals “tens of millions,” $1.5B pipeline (not bookings). Partners also HPE, Lenovo, Advantech, Axiomtek. Not a training GPU — production inference inside the power plant you already own.
Why it matters
This is a PCIe swap in the NVR/analytics server, not a locked AI camera. Spec Europa/Metis + Voyager, swap YOLO/VLM later. Accelerated compute beats locked “AI cameras”: swap models vs a 36–84 month hardware refresh.
Source
Reuters; Network World - 7
Key point
Factory VLA inference is latency-bound. Repeated station jobs share not just pixels/actions but internal model states. rMuscle: dual-phase cache — Context Cache reuses visual tokens, Action Cache reuses neuron activations. Online recompute + sliding-window retrieve + mask sharing across denoising steps keep the cache small. 1.29–1.42× on RTX 4090 and Jetson Thor (LIBERO, RoboTwin, physical manip) with no success-rate drop on real robots.
Why it matters
Plant/AMR cells will ask why the VLA is slow on the box you already sold. Cache the repeated station, don’t buy a locked camera brain. Accelerated compute beats locked “AI cameras”: swap models vs a 36–84 month hardware refresh.
- 8
Key point
MOT still dies when the target leaves the frame. Pipeline: YOLOv11n detect → Kalman during hide → occlusion-aware ReID on reappear. Six ReID nets compared; OAMN won. Vs OccluTrack on OVIS: +18.1% MOTA, +25.1% IDF1, −12.8% ID switches. Custom military/surveillance set: MOTA 0.734, IDF1 0.729. Aimed at defense/surveillance that cannot drop a track behind a truck.
Why it matters
Perimeter/plant MOT RFPs fail on occlusion, not on daytime mAP. Spec ReID-through-hide, not another locked “AI camera” that forgets the person behind the forklift.