IVS Briefs
IVS / Edge AI Research Brief
Overall assessment
Signal level
High on a capital-city Flock non-renew with an LPR-swap door (Jackson), a state ALPR pause that expires today (Rhode Island), and an NVR-class 64-ch decode SoC (Axera) the same Monday.
Notable trends
- Week catchup is chief/state policy, not another six-city terminate list — Jackson keeps an RTCC LPR door; RI’s freeze lifts today with a State Police memo still in edits.
- Tonight’s live votes are Quincy 49 and Sherman 11; Troy’s override is Thursday.
- Edge this window is a 64-ch decode Axera module, not Hailo/Dragonwing.
- Monday cs.CV dump is 106 new; Lab is LiDAR+thermal person-detect, RGB–IR MOT, and orthogonal multi-modal ReID.
Gaps
No 48h/week VMS SKU (Genetec/Milestone/Verkada/Axis silent; Genetec SaaS Axis/Bosch/Hanwha item is a 2024 reprint). No Hailo / DeepX / Qualcomm module SKU (Geniatech AIM-M already 10-14; Spotter already 10-15; Orange Pi O1 already 10-16). No World Labs / NVIDIA Metropolis post (AMD–World Labs already 09-29; RealSense D555+Thor is an Aug 2025 reprint). Norwell 2-cam Sept 30 pull left unnumbered (Patriot Ledger recap).
Mix counts
LPR 1 / VMS 0 / Edge 1 / Policy 4 / Lab 3. LPR+Policy city/contract = 5/9 = 56%.
Now vs On deck
Now 2 (Jackson, RI pause) · On deck 7 (Quincy tonight, Sherman tonight, Troy Oct 22, Axera, three papers).
Signals (9)
- 1
Jackson, MS: Chief Brackney will not renew Flock; cameras stay until expire; RTCC still wants some LPR
Key point
Jackson Police Chief RaShall Brackney told the Parks and Public Safety Committee the city will not renew Flock. Quote: “We are not renewing Flock.” Council approved the contract September 2024; WAPT: dozens of cameras since 2024. The chief administrative officer could not give an expire date; Flock removes the hardware when the term ends. Brackney will still “look at license plate reader systems” for the Real-Time Command Center, “with stringent policies.” She cited 50+ officer-misuse cases nationally and mandatory sharing she refused in Charlottesville. Council President Vernon Hartley (who voted yes in 2024) wants a “good discussion” at renewal, not a kill. Ward 4 Brian Grizzell wants a policy, not a pull — “cameras have helped solve a lot of crimes.”
Why it matters
This is a chief-level Flock exit with an LPR-swap door. Hardware stays until the unknown expire date. Bid a local-only plate stack that plugs into the existing RTCC, not a Flock reboot and not a dark city.
Source
WLBT; WAPT - 2
Rhode Island 60-day ALPR pause expires today; State Police policy in “final edits”; ACLU / Public Defender left out
Key point
Gov. Dan McKee’s Aug. 18 60-day freeze on new ALPR installs ends today. Lt. Col. Erik Yanyar: State Police were making final edits to a statewide policy as of Thursday. McKee told DPS (Col. Darnell Weaver) to work with the Police Chiefs Association and League of Cities and Towns. ACLU of RI, the Public Defender, AMOR, and Planned Parenthood of Southern New England’s Oct. 9 letter says the working group never asked the public. Chiefs’ Sidney Wordell: “nothing new” in that letter; he would not limit ALPRs to violent felonies (property crime). Already off Flock: South Kingstown, Narragansett, Glocester, North Kingstown. Tiverton approved 4 cameras that are not yet installed. Woonsocket first-passage ordinance on 15 cameras (needs a second vote). Providence’s ban ordinance is already numbered (10-16 referral).
Why it matters
Today is a state-policy day, not a city pull. Tiverton’s four and any new municipal install wait on Weaver’s memo. Do not bid a statewide Flock expansion until the policy is public; local-only retain/share language is what the ACLU letter is asking every chief to copy.
- 3
Quincy, MA tonight: Res. 2026-143 asks the mayor to terminate 49 Flock ALPRs and cover until removal
Key point
Councillors Ziqiang Susan Yuan (at-large) and Virginia M. Ryan (Ward 4) introduced Order 2026-143 for tonight’s council. It calls on the administration to terminate all Flock contracts, remove or cover the cameras, and stop collection during unwind; a companion ordinance would require a public hearing and council vote before any future surveillance buy. PD runs 49 ALPRs plus 1 Condor dump camera (Chief Mark Kennedy). Flock since 2020. Oversight hearing Oct. 6: Kennedy said 689,000 plates captured / 76 searches in 30 days, 48 officers authorized; Norfolk DA Michael Morrissey asked council not to drop the tool. This vote does not itself pull the cameras — it directs the mayor.
Why it matters
Forty-nine live ALPRs is a South Shore-scale unwind if the mayor complies. Hardware stays until the administration acts. Do not quote a swap until tonight’s vote and the mayor’s written response.
- 4
Sherman, TX tonight: Res. 7682 would not renew Flock, bag 11 cameras, and lock a public hearing before any future ALPR
Key point
Sherman City Council takes Resolution 7682 tonight, directing the city manager not to renew or extend Flock Group Inc. Mayor Shawn Teamann put it on the agenda after constituent pushback. If it passes, staff alerts the manager to remove the city’s 11 Flock cameras; Teamann: the ordinance would be effective immediately, removal takes “a little bit of time.” A future ALPR install would need a public hearing first. Teamann drafted it as a Grayson County blueprint other cities can copy.
Why it matters
A Texas no-renew + hearing lock after the Comptroller grant ban. If it passes, do not bid Flock or a quiet Axon swap in Sherman — any plate stack comes back through a noticed hearing.
Source
KXII - 5
Key point
Friday’s brief numbered Mayor Carmella Mantello’s veto of Local Law 3 (72-hour retain, felony-related search, court-or-DA share). New this window: Council President Sue Steele said in writing the all-Democrat council “will override the veto on October 22.” Spectrum: she called the veto message “factually incorrect.” Hardware stays. The live fight is whether 72-hour retain becomes city law next Thursday.
Why it matters
Calendar lock, not a pull. Do not bid a rewrite or a bag-and-tag until the Oct. 22 override lands.
- 6
Axera AX8860P: 16× A55, 64-ch decode, 128/256 TOPS, 78×78 mm, ~40 W; Sipeed MaixStation <$2k / <$4k, Kickstarter months out
Key point
Axera AX8860P is a 16-core Cortex-A55 SoC, 6 nm, 78 × 78 mm module: NPU up to 128 TOPS / 256 TOPS (FP4), 256-bit LPDDR5/5X up to 128 GB at 273 GB/s, PCIe 5.0 plus AXLink C2C, ~40 W under load. VPU: 64-channel decode, 16-channel encode. Sipeed MaixStation mini PC: 110 × 110 × 30 mm, 32 GB or 128 GB, USB-C or 10GbE SFP+, 65 W PD; claimed 32 GB under $2,000 / 128 GB under $4,000. Not for sale; Kickstarter “in a few months,” $30 reservation for $100 off. Radxa “Pocket Supercomputer” still undescribed. CNX: Sipeed’s Qwen numbers vs DGX Spark ($6,950) are close; vendor claims dense 256 TOPS is more usable than NVIDIA’s sparse 1,000 TOPS slide.
Why it matters
64-channel decode on a 40 W module is an NVR-class experiment, not a locked ALPR camera. Accelerated compute beats locked “AI cameras” (model swap vs 36–84 month refresh) — quote it only after Sipeed actually ships a board that boots.
Source
CNX Software - 7
TriSAR-3D / PaintFusion: LiDAR + RGB + thermal 3D person detect; 65.4 AP@50, 51.6 ms; lying-pose models collapse without lying training
Key point
TriSAR-3D is a search-and-rescue person-detect set on 128-channel LiDAR + RGB + thermal, indoor and outdoor, including lying and occluded poses automotive sets skip. PaintFusion (PointPainting evidence into a simple MLP) hits 65.4 AP@50 vs CenterPoint LiDAR-only 53.4, BEVFusion RGB+LiDAR 60.5, Tri-BEVFusion 60.7, at 51.6 ms/frame vs BEVFusion 101 ms. Train without lying people and lying-pose AP@50 goes to ~0 on every detector they tried. Dataset and code on GitHub.
Why it matters
Perimeter / parking-garage “person down” is a pose problem, not a TOPS problem. If the RFP is occupancy or man-down, ask whether the training set has lying bodies and whether thermal is in the fuse — a road-pedestrian checkpoint will miss them.
Source
arXiv:2610.22509 - 8
Key point
RCA-MOT keeps RGB-only, IR-only, and fused ID descriptors as separate appearance states. A training-free reliability score per branch (1) picks which ID bank to trust, (2) calibrates borderline ReID accepts, (3) throttles how hard a match writes into appearance memory — so a washed-out night RGB frame cannot poison the IR track. New SOTA on VT-MOT and UniRTL: HOTA +5.402 / +6.075, IDF1 +11.549 / +12.789 vs prior best. IEEE TMM submission. Code promised.
Why it matters
Dual-imager (RGB + thermal) sites fail when fusion averages a dead modality into the ID. Ask the VMS/analytics vendor for per-camera reliability, not a single fused embedding, before you buy a night-time ReID add-on.
Source
arXiv:2610.23125 - 9
ODSS-Net: orthogonal shared-structure / modality-specific ReID so alignment does not erase the extra camera
Key point
ODSS-Net (ICASSP 2027 submission) splits multi-modal object ReID into (1) a shared structural branch that suppresses appearance that only one sensor sees, (2) orthogonal specific branches that drop the shared component by geometric projection, (3) a sample-adaptive mixture of experts that uses the shared structure as routing context to fuse the leftovers. Claim: ordinary alignment entangles shared geometry with sensor-specific cues, and ordinary fusion then wastes the extra camera. Three unnamed benchmarks in the abstract.
Why it matters
RGB+IR+LiDAR ReID RFPs fail when the integrator is sold “we aligned the features.” Ask whether identity-specific cues are projected off the shared structure, or you are paying for a second sensor that the embedder already threw away.
Source
arXiv:2610.22922