IVS Briefs
IVS / Edge AI Research Brief
Signals (9)
- 1
Key point
City Council 5–2 Monday night (Third Ward Jacque Sample and Sixth Ward Betsy Peters opposed) stopped automatic renewal of Columbia’s Flock Safety contracts after a two-hour public hearing. Mayor Barbara Buffaloe: “Council voted not to renew our Flock Safety contract when it expires in 2027.” The vote does not immediately terminate. City staff: three active Flock contracts — two expire March 2027, Utilities September 2027. Sample said she understood the night as discussion, not a final vote. Buffaloe: she still wants public-safety tech, but “recent red flags” and Flock’s “long-term viability” mean a more reliable partner.
Why it matters
A 2027 hole, not a dark city. Quote March / September 2027 expirations and bid a local VMS/LPR against the three-contract stack; do not certify “off” from Monday’s 5–2.
Source
KOMU - 2
Key point
Informational session on CRPD switching 70 ALPRs from Flock (contract ends February) to Axon. Staff slide: Flock $225,000/yr vs Axon $98,415/yr, save >$125,000; no expansion beyond 70; reuse city infrastructure. Mayor Tiffany O’Donnell: the city has not voted the ALPR expansion; the swap of 50 of 70 cameras is on the next council meeting. Capt. Cody Estling: no ICE share, local reciprocal only. Public pressed why feedback came after the city-manager email (September) already named Axon; Councilor Dale Todd told critics to leave.
Why it matters
An Axon in-place swap at ~$1,406/cam-yr vs Flock ~$3,214, still unvoted. Show up at the next CR council with a camera-agnostic LPR quote; do not treat the September email as the award.
Source
CBS2 Iowa - 3
Flock’s policy lead: “strong legislation state by state”; 120k cameras, 7-day default, case numbers by year-end
Key point
Flock senior director of policy Andrea Korb (R Street webinar): the company expects an “onslaught” of 2027 bills and supports “strong legislation state by state.” Pluribus: >120,000 cameras in 49 states; about half of states have ALPR statutes, ≥6 new laws this year. Flock valued $8.3B in August; ≥$2M lobbying and lobbyists in ≥34 states (OpenSecrets). Vendor moves already announced: default retain cut to 7 days (August), misuse alerts, case-number required searches by year-end. Korb: vendor changes are not enough; “unfettered access to all vehicles and all cameras for any reason” should end. She called a warrant-to-query rule “administratively burdensome” with a “massive efficacy reduction”; Libertas’s Emily Grimmius is drafting Utah 2027 warrant + council-approval + audit language and cited Minnesota’s third-party audit.
Why it matters
2027 RFPs will ask for warrant/audit/case-number language the vendor now says it wants. Price retain, query gates, and third-party audit as line items, and do not bid “unfettered share” into a state that is about to close it.
Source
Arizona Capitol Times - 4
NY Assembly ALPR hearing: 10 witnesses, no floor vote; Orange County unanimous vendor-ALPR ban last week
Key point
Assembly Consumer Protection (chair Nily Rosic, D-Queens) took 10 witnesses Monday on ALPR use, retain, and share. Suffolk DA Raymond Tierney: keep the tool, add safeguards. Assemblyman Jonathan Jacobson (D-Newburgh/Poughkeepsie): at minimum apply existing speed/school-zone camera restrictions. AG’s Tyler Nims: longer retain paints a picture of millions of non-offenders. STOP’s Corrine Worthington: ICE / protester / out-of-state reproductive tracking. Rosic: New York “lack[s] universal standards”; many bills are in the hopper; no floor vote was scheduled. Mid Hudson: Orange County Legislature unanimously voted last week to ban commercial/vendor-linked ALPRs on county roads (Ulster EO and Dutchess deactivation already numbered).
Why it matters
Statewide ALPR language is being written in Albany while Hudson Valley counties already banned or pulled vendor systems. Do not bid a Flock-class share network into NY without a county-road / retain / warrant read of the pending Assembly package.
Source
Mid Hudson News - 5
Key point
HB 331 (Rep. John Bucy): government entities statewide may not use ALPRs, vendor-agnostic; private right of action for actual damages or $1,000 statutory plus fees. HB 847 (Rep. Brian Harrison): warrant or court order to use ALPRs; data only for specified violent crimes; prompt destroy unless the plate is in a qualifying case; annual public reports. HB 763 (Rep. Mary González): allow felony / missing-person / stolen-vehicle uses; generally destroy plate data in 48 hours except evidence/hold/warrant. Houston Police Officers’ Union president Douglas Griffith wants the cameras kept with misuse penalties. Conroe resident James Fletes: ~1,700 signatures toward a May surveillance ordinance, still unverified. All three are proposals; House, Senate, and the governor still required.
Why it matters
Texas is no longer only city bag-and-remove. Map live ALPRs against ban vs warrant vs 48-hour and do not write a multi-year Flock/Axon/Motorola term that a 2027 session can void.
Source
Click2Houston - 6
Key point
Tessera Defense and Homeland Security (NYSE American: HLSQ) launched a commercial perimeter stack: DFSL LADAR (laser detect-and-range, no RF radar) fused with Zorronet LYBA C2. Detect at about 30, 60, and 100 m; on hit, LYBA slews a compatible day/night or thermal PTZ, then classifies person / group / person+dog / vehicle / visible weapon. Camera-agnostic commercial PTZ, not a locked imager. On-prem and air-gapped; models from a library plus third-party/open-source, validated before deploy; software/model updates without replacing the sensor. Aimed at defense, energy, logistics, and RF-restricted sites. 8-K flags a going-concern qualification and NYSE American continued-listing risk. No street price in the launch.
Why it matters
A shipping short-range 3D detect → PTZ-verify box you can hang on an existing thermal/PTZ, with model-swap without a sensor rip. Diligence HLSQ listing risk before a multi-site PO. Accelerated compute beats locked “AI cameras” (swap models vs a 36–84 month refresh).
- 7
PTC-IoU: CPU-side detection-quality estimator on Jetson Orin Nano — 4.7× faster, −95% estimator energy
Key point
Wei Geng et al. (TUM): edge video analytics has no runtime ground truth, and detector confidence misses both FN and bad boxes, so a quality estimator must decide which frames to escalate to a cloud model. Neural estimators compete for the same GPU as detection. PTC-IoU (Predictive Temporal Consistency IoU) moves scoring to CPU by compacting box-history. On Jetson Orin Nano: estimator latency 10.3 → 2.2 ms (4.7×), pipelined throughput +13.6%, estimator energy −95%, competitive final detection quality on multiple datasets.
Why it matters
Ask edge-analytics vendors how they gate cloud fallback without a second GPU network. A CPU IoU history is a cheaper SLA than “every frame to the VLM.”
Source
arXiv:2611.12405 - 8
Key point
Wei Zhang et al.: unsupervised person ReID usually locks to one backbone. MDA (Mamba-augmented multi-architecture domain adaptation) wraps Mamba as an MDA operator, then a framework that is not tied to a single insertion point. On CUHK03 → Market-1501, CNN backbone +10.2% mAP; Transformer backbone +19.8% mAP. Authors: a single deployment config already beats several SOTA unsupervised ReID methods.
Why it matters
Cross-camera ReID demos die on a new site’s cameras. Ask for unsupervised domain-adapt mAP on your cameras, on both CNN and Transformer SKUs, before a campus or downtown MTMC.
Source
arXiv:2611.11836 - 9
LO-World: joint 4D occupancy + LiDAR forecast in one BEV stream; occupancy IoU +10.9%, LiDAR CD −30.1%, >25 FPS
Key point
Zhimin Liao et al.: occupancy (dense volumetric semantics) and LiDAR (metric geometry) are usually forecast in separate or cascaded world models. LO-World does both in one BEV latent: hybrid range-voxel autoencoder + flow-matching generator with cached history. Occupancy forecasting IoU +10.9%, LiDAR Chamfer Distance −30.1% vs prior SOTA; >25 FPS on one RTX 4090. Code posted.
Why it matters
Perimeter and curb occupancy is moving from 2D boxes to 4D LiDAR+occupancy. Ask physical-AI / LiDAR vendors for a joint occupancy-and-points FPS number, not two pipelines stitched in the VMS.
Source
arXiv:2611.11912