IVS Briefs
IVS / Edge AI Research Brief
Signals (9)
- 1
Key point
Reno City Council approved an expanded Axon Enterprise contract 5–2 after hours of public comment. No: Mayor Hillary Schieve and Councilwoman Meghan Ebert. News 4: the updated agreement runs eight years and comes in just under $22 million, adding ALPRs, drone technology, and AI report drafting on top of existing body-worn cameras, TASER devices, and evidence storage. This Is Reno’s Oct. 15 staff-report writeup: $21,776,231.26, Jan. 1, 2027–Dec. 31, 2034; more than 200 ALPRs — 136 vehicle-mounted plus 50 solar Outpost readers ($312k fixed). Council conditions quoted by News 4: ALPR data not tied to a police investigation deleted after 14 days; the city keeps control of its data; RPD quarterly audits; Chief Corey Solferino reports to council every three months; any new surveillance outside the agreement needs council approval; the city can end the deal with or without cause on 30 days’ written notice. The city had already told Flock it will not auto-renew (current term ends March 17, 2027).
Why it matters
This is a Flock-to-Axon swap with a retain/audit rider, not a plate-SaaS exit. Bid Axon Outpost / in-car ALPR plus the 14-day delete and quarterly-audit language; do not quote Flock down until March 2027.
- 2
Appleton 11–1 Res 16-R-26: city shall not acquire/operate ALPRs except parking, access, weigh; 30-day data
Key point
Appleton Common Council adopted Res #16-R-26 11–1. Sole no: Council Member Brad Firkus (AppletonToday). Sponsor Adrian Stancil-Martin (District 11). The city shall not acquire, install, operate, use, or fund ALPRs except parking enforcement, access control to non-public restricted areas, and commercial-vehicle weigh-station regulation. Data under those carve-outs may be generated, stored, or shared only for those purposes and generally cannot be preserved more than 30 days unless a longer period is required by state or other applicable law. 14 speakers, all in favor. Appleton already stopped Flock over the summer; the resolution is vendor-class, not a Flock-only kill. Alderpersons are still drafting a separate mass-surveillance guardrail. Christopher Croatt flipped from a committee no after the amendment.
Why it matters
Appleton is a vendor-class ALPR lock with parking/access/weigh carve-outs. Do not bid citywide plate SaaS; a parking or gate-access stack with 30-day local retain is the only live path until council reopens the topic.
Source
NBC 26; AppletonToday - 3
Key point
The Budget and Government Efficiency Committee voted 4–0 (Henry Foster III, Kent Lee, Vivian Moreno, Sean Elo-Rivera) to rescind city approval of ALPRs and end the Flock contract. Councilmember Stephen Whitburn flipped last week, so a full-council majority now exists on paper. Mayor Todd Gloria supports keeping the cameras and is expected to veto; override needs 6 of 9. UT: 500 streetlight surveillance + ALPR units citywide; the package is Ubiquia hardware with Flock software at $2 million/year; SDPD retains data 30 days unless used in an investigation. Chief Scott Wahl: ALPRs helped in more than 900 cases. Full council is expected before year-end under the 2023 surveillance-ordinance annual review. This is a committee lock, not a take-down.
Why it matters
San Diego is a calendar + veto fight, not cameras off. Do not quote 500 readers dark; watch the full-council date and the 6-vote override math. Ubiquia streetlight hardware without Flock software is the residual bid if ALPR is stripped.
- 4
Anderson County, SC 7–0 second reading of Ord. 2026-054; third reading November; 10 cams already canceled
Key point
Anderson County Council unanimously approved second reading of Ordinance 2026-054 (packet: 7–0; Greg Elgin moved, Jimmy Davis seconded) to bar county funds, property, IT, and personnel from Flock camera systems, ALPRs, software, databases, and subscriptions. Councilman Chris Sullivan (WYFF): one more third reading in November; the county has already canceled subscriptions for 10 cameras at parks and boat ramps, Flock just has to pick up the hardware, and the county saves $30,000–$40,000. Sheriff Chad McBride supports. Elgin on the dais: only one of the 10 was on a public roadway. City of Anderson, HOAs, and businesses (including Lowe’s) keep their own readers; nine municipalities can still buy. Do not treat second reading as law.
Why it matters
County cameras are already a non-renew / pickup, and the ordinance is a November calendar lock. Do not bid county-ROW Flock; City of Anderson and private property remain the live channel.
- 5
Key point
Raytron’s sponsored CNX post details the SE5 1280 uncooled thermal core: 1280×1024 VOx, 12 μm pitch, 8–14 μm LWIR, NETD ≤40 mK, ≤29 × 29 × 15.7 mm, <26 g, thermal video up to 60 Hz and radiometry up to 30 Hz. The onboard LY300 ISP (32-bit MCU, up to 1920×1080@30, 16 MB flash / 128 MB LPDDR2) does NUC, dead-pixel correction, spatial/temporal NR, WDR, 8× digital zoom, isotherms, and OSD, with DVP, MIPI, USB 2.0, VOSPI, UART, I²C, SPI. Raytron’s own brochure lists security, handheld, and night-vision OEM builds. This is a VISION 2026-shown module getting first CNX coverage in the 48h window, not a Jetson/Hailo NPU SKU.
Why it matters
Quote SE5 1280 when an MSP needs a swap-in 1.3 MP thermal core for perimeter or counter-UAV boxes instead of a locked dual-spectrum camera with a 36–84 month refresh.
Source
CNX Software - 6
Conditional Imaging: camera optical state per frame conceals the scene until authorized reconstruction
Key point
Hengyu Lian, Debabrata Mandal, Anzhou Wen, Yujie Wang, Xi Peng, and Praneeth Chakravarthula argue that static optics cannot be both private and later-recoverable: anything left in the measurement is also available to an observer. Conditional imaging changes the camera’s optical state every frame so the measurement alone conceals the scene; reconstruction needs the matching physical system state. They built a prototype and evaluated it on a captured faces dataset against direct visualization, computational inversion, learned reconstruction, and identity recognition. Authorized reconstructions preserve identity; observers without the system state do not get usable visuals.
Why it matters
This is privacy at capture, not a 30-day cloud retain policy. Watch it as a camera-architecture option for jurisdictions that want recoverable evidence without a live identifiable stream.
Source
arXiv:2610.28835 - 7
Key point
Rajarshi Bhattacharya, Aryan Shukla, Shakeeb Murtaza, Christian Desrosiers, Jose Dolz, Maguelonne Heritier, and Eric Granger propose PRAISE (Prompt-guided Attribute-based Inference-time Similarity Enhancement) for VLM person ReID. Templated attribute prompts (clothing color, carried objects) are biased if used as scores. PRAISE instead builds a query-dependent low-rank Attribute-Conditioned Subspace from a transferable prompt bank and reshapes query–gallery affinities at test time with no identity or attribute labels. It improves both image-to-image and text-to-image ReID backbones versus generic and ReID-specific test-time adaptation.
Why it matters
This is gallery search without retraining — clothing/object attributes as a test-time rerank, which is the path MSPs get asked for when a locked camera ReID model drifts across sites.
Source
arXiv:2610.27944 - 8
Key point
Zheng Qin, Junhua Xi, Yan Huang, Jiahong Lai, Anwen Huang, Qiong Li, Guangda Zhang, and Kai Xu train a single Register Anything Model (RAM) for generalizable 3D point-cloud registration. Geometry-aware voxelize-and-normalize maps indoor/outdoor/object clouds into a shared geometric space; a hierarchical coarse-to-fine geometric transformer scores local consistency instead of global all-pairs attention. On 12 benchmarks RAM beats domain-specific specialists on inlier ratio and registration recall. Code: github.com/qinzheng93/RegisterAnything.
Why it matters
Multi-sensor perimeter and occupancy stacks fail first on cross-LiDAR registration. A single 6-DoF model that survives indoor-to-outdoor density shifts is the Lab path into a camera+LiDAR box.
Source
arXiv:2610.28821 - 9
Key point
Alberto De Zan, Denis Tavaris, Gian Luca Foresti, and Ivan Scagnetto compare early (image-level RGB+IR into one detector) vs late (fuse bounding boxes from two pipelines) fusion for UAV detection. Tests cover low/high brightness, sky and ground backgrounds, 5 / 25 / 60 m, and 1× / 2× / 4× zoom. Early fusion wins raw detection because thermal is in the training image; late fusion is the fallback when sensors or frames are misaligned. Practical recommendations are in the paper; this arXiv copy includes reviewer revisions that missed the IEEE Access version.
Why it matters
Dual-spectrum perimeter UAV detection is an RFP item this quarter. Quote early-fusion when RGB and IR are hardware-aligned; quote late-fusion when an MSP is bolting a thermal module (see item 5) onto an existing visible camera.
Source
arXiv:2610.29828