IVS Briefs
IVS / Edge AI Research Brief
Signals (9)
- 1
Key point
Lafayette City Council voted unanimously to remove the city’s Flock cameras after the Oct. 6 first reading. A city spokesperson: Lafayette will immediately end all use of Flock Safety technology, begin early termination of the contract, cover all 30 cameras under the city contract until Flock removes them, and staff will not contract with any other ALPR without a council directive. Flock data was already suspended after the Oct. 6 meeting. The city also flagged that not every ALPR near Lafayette is on the city contract — corporate and neighboring-municipality readers stay. Flock has been in Lafayette since 2022 ($77,500/yr).
Why it matters
This is a bag-and-tag plus a vendor-class lock, not a Flock-to-Axon swap. Do not bid another plate SaaS into Lafayette unless council puts it on the dais; nearby non-city ALPRs are still live.
- 2
St. Johns County, FL 3–2: revoke county-road LPR permits; SJSO says 89 county-ROW cameras, no disable date
Key point
St. Johns County commissioners voted 3–2 to revoke use permits for ALPRs on county rights-of-way — the last Northeast Florida holdout after FDOT pulled state-road permits. Yes: Sarah Arnold, Ann Taylor, Christian Whitehurst. No: Krista Joseph and Chair Clay Murphy. Sheriff Rob Hardwick opposed; SJSO statement quoted by News4JAX: the vote requires removal of 89 LPRs on county ROW, with no effective date to disable; SJSO already pulled 84 on state ROW in September; HOA / private-property readers stay and can still share with SJSO. CID director Howard Cole: the office now shares with zero outside agencies (down from 2,600) and found one personal-use case in an audit. Hardwick 2026 stats: 29 stolen vehicles, 35 wanted, 35 missing. Whitehurst would bring cameras back if the Legislature adds guardrails.
Why it matters
Florida county-ROW is now a permit kill, not a city terminate. HOA and CDD cameras remain the live channel into SJSO. Do not quote a pull date until the commission sets one.
- 3
Key point
After discussion, the Florence County Board voted unanimously to discontinue Flock. The county has two active cameras; they will be turned off and bags put over them until they are removed. Chair Chad Hedmark. Same night the board also unanimously passed an 18-month data-center / BESS moratorium so zoning can rewrite the 20-year Comprehensive Plan. WLUC did not publish a named no-vote or an unwind day-count.
Why it matters
Small-county bag-and-tag is live this week. Do not bid a swap into Florence unless the board puts a replacement on the agenda; the hardware is covered until Flock pulls it.
Source
WLUC / TV6 - 4
Rhode Island: McKee extends the ALPR pause indefinitely; State Police model policy is 30-day retain, not statute
Key point
Gov. Dan McKee left binding ALPR safeguards to the General Assembly and extended the new-install pause indefinitely. Col. Darnell Weaver delivered a statewide model policy and data-sharing MOU template: delete plate data within 30 days unless preserved for an investigation; bar use for immigration enforcement, traffic tickets, outstanding fines, civil infractions, First Amendment, discrimination, or reproductive care; require a public transparency portal. ACLU of Rhode Island flagged “fundamental concerns”; advocates had asked for 7-day retain (Flock’s August default). ACLU estimates at least 280 cameras statewide. House and Senate leaders had not reviewed the draft; they expect the 2027 session. The pause itself is not a take-down of cameras already in the field.
Why it matters
Rhode Island is still a do-not-ship state for new ALPRs. Existing Flock stays until a city terminates or the Assembly writes a statute; bid only local-only retain/share stacks that can survive a 7- vs 30-day fight.
Source
Rhode Island Current - 5
Key point
A motion to suspend council rules and immediately consider the resolution telling Mayor Wade Kapszukiewicz’s administration to terminate Flock, deactivate or remove equipment, and delete data failed 7–5 — two votes short of the nine needed to skip the two-reading rule. Yes to suspend: Mac Driscoll, Theresa Gadus, John Hobbs III, Brittany Jones, Nick Komives, Erin Kramer, Sam Melden. Hold: Theresa Morris, George Sarantou, Vanice Williams, Adam Martinez, Cerssandra McPherson. The resolution returns Nov. 4. Williams wants the independent Flock audit first (auditor John Rivalsky; 30-day ask, not a hard deadline). Council agenda review is Oct. 27. The resolution is a mayor-urge, not a self-executing terminate.
Why it matters
Toledo is a calendar lock, not a pull. Do not quote cameras down; Nov. 4 is the next floor vote and the audit is still open.
Source
The Blade - 6
Key point
Aeva (Nasdaq: AEVA) and Sensys Gatso announced regulatory certification of Aeva’s FMCW 4D LiDAR as a secondary speed-measurement device inside Sensys Gatso’s automated speed-enforcement stack for public-road deployment in Australia. Atlas independently measures vehicle position and velocity across multiple lanes and is integrated with Sensys Gatso radar + high-resolution cameras. Aeva: a single sensor can monitor up to six lanes, approaching and receding. The pair called it a first-of-its-kind certification for 4D LiDAR inside an automated traffic-enforcement system and said it supports commercial rollout in Australia plus other markets.
Why it matters
This is a LiDAR + camera enforcement box with a legal secondary-speed path, not a city ALPR SaaS. Quote Atlas/Sensys Gatso when an MSP is asked for multi-sensor speed evidence that a locked ALPR camera cannot provide.
Source
Aeva - 7
CodecMotion-OAD: decoder motion vectors replace optical flow for streaming action detection; 11.48× CPU
Key point
Sun, Zheng, Xia, and Wu (CodecMotion-OAD) apply codec motion vectors already available at decode time to frame-level online action detection, dropping a separate optical-flow pass. Two-stage training aligns MV features to a pretrained flow interface, then trains RGB+MV detectors from a frozen RGB+Flow teacher. On THUMOS’14 with MiniROAD: 68.3 mAP vs 70.9 mAP for matched-distillation RGB+Flow, while decoder-to-prediction time falls 11.48× on CPU and 2.04× on GPU. Authors call it the first use of codec MVs for frame-level OAD.
Why it matters
Streaming VAD/OAD on the cameras a city already owns is cheaper if you skip dense flow. Bid this pattern when a council wants “analytics on existing CCTV” after a Flock pull.
Source
arXiv:2610.26915 - 8
Key point
Song and An add a training-only foreground-attraction loss to GaussianFlowOcc-style camera-based occupancy: back-project 2D foreground pseudo-labels with pseudo-depth and pull nearby Gaussians toward those 3D points. Inference is unchanged. On Occ3D-nuScenes, Gaussian density on foreground voxels rises 1.62× → 2.40× vs background; mean foreground IoU 6.29 → 6.96 over three seeds. Mean mIoU 13.53 → 13.82 sits inside baseline seed noise. Ablations: pseudo-label targets beat ground-truth voxels; acting on Gaussian positions alone is not enough.
Why it matters
Camera-only occupancy is the path to people/vehicle voxels without a LiDAR truck. Quote it when an MSP is asked for 3D occupancy on existing CCTV, not another plate pole.
Source
arXiv:2610.25631 - 9
Jetson AGX Orin depth bake-off: NVIDIA ESS 0.70% AbsRel @ 8.6 ms; monocular models scatter 17–46% of edges
Key point
Jiang’s 2026 Orin benchmark runs nine depth families as TensorRT FP16 / native CUDA on Jetson AGX Orin (up to 5.7 MP): Depth Anything V2/3, Depth Pro, MoGe-2, Metric3D v2, UniDepth V2, libSGM, NVIDIA ESS, Fast-FoundationStereo. Stereo wins. At 30 fps, NVIDIA ESS hits 0.70% AbsRel in 8.6 ms and libSGM 0.73% in 25 ms; Fast-FoundationStereo reaches 0.33% with more compute. Best monocular with true focal length is ~12% (Depth Anything 3 Metric-L, 33 ms); 3.8–4.0% only with an oracle per-image scale. At depth discontinuities, monocular models scatter 17–46% of edge points between foreground and background vs 2–9% learned stereo and 1% full-res libSGM. Code: github.com/nodarhub/orin-depth-benchmark.
Why it matters
Perimeter occupancy and people-counting on Orin should ship stereo ESS, not a monocular foundation-depth demo. Accelerated compute beats locked “AI cameras” (model swap vs 36–84 month refresh).
Source
arXiv:2610.27199