IVS Briefs
IVS / Edge AI Research Brief
Overall assessment
Signal level
High on a first-in-Arizona city-code ALPR ban, a Bloomington unanimous third-party-share lock, and a Downey Motorola expansion that went the other way the same week.
Notable trends
- Overnight is ban-in-code (Tempe, Bloomington) plus a veto fight (Troy) plus a Motorola grow (Downey) — not another Flock terminate.
- Tempe and Bloomington both carve out parking/CCTV/body-worn and leave campus/state ALPRs live.
- Edge this window is a $195 SOPHGO 16-ch decode SBC, not Hailo/Dragonwing.
- Friday cs.CV dump is 135 new; Lab is weight-memory streaming, sports MOT, and FPGA occupancy with a real tail bound.
Gaps
No 48h VMS SKU (VIVOTEK/March Networks page did not resolve; Genetec/Milestone silent). No Hailo / DeepX / Qualcomm module SKU (Orange Pi is SOPHGO; Lantronix Spotter already 10-15). No World Labs / NVIDIA Metropolis post (Nemotron 3.5 Super VL hit a checkpoint; RTX Spark recaps already numbered 10-12). Jackson, MS Clarion Ledger “chief to end Flock” is paywalled — skipped. Fullerton OCR “cameras go dark” is the already-numbered 10-09 pause taking effect — skipped. League City still paywalled.
Mix counts
LPR 1 / VMS 0 / Edge 1 / Policy 4 / Lab 3. LPR+Policy city/contract = 5/9 = 56%.
Now vs On deck
Now 5 (items 1–4, 6) · On deck 4 (Providence referral, MaLoW, ARC-SORT, FALCON).
Signals (9)
- 1
Tempe 6–1: first Arizona city-code ALPR ban; 33 Flock already dark; 30-day effective; ASU / parking carved out
Key point
Tempe City Council voted 6–1 (no: Vice Mayor Arlene Chin) to put an ALPR ban in city code — first in Arizona. It bars the city and PD from purchasing, using, funding, or accessing ALPR systems and data except by warrant / subpoena / court order, unsolicited third-party data tied to a specific case, or task-force review of another agency’s lawful hit that Tempe did not request. Existing ALPR data must be destroyed unless evidence or legal hold; employees who violate face termination and civil/criminal penalties. Effective in 30 days; only a future council vote can repeal. August already canceled Flock and bagged 33 city cameras. Parking ALPRs stay if they do not feed a searchable location database. ASU’s 20 campus Flock cameras are untouched. RTOC’s hundreds of non-plate cameras stay. Chief Ken McCoy and the union opposed; Councilmember Bobby Nichols sponsored. Chin: a wholesale ban is “short-sighted.”
Why it matters
This is a code lock, not a Flock-only pull. An Axon/Motorola swap still needs a repeal. Bid parking, campus (ASU), or non-plate RTOC — not a citywide ALPR reboot.
Source
AZFamily; KTAR; FOX 10 - 2
Bloomington, IN unanimous: ban ALPRs and third-party-share surveillance cameras; 11 Flock already down
Key point
Bloomington City Council voted unanimously to ban ALPRs and any surveillance camera whose footage can be shared with a vendor or another government. Parks CCTV and body-worn cameras are carved out. Goes to Mayor Kerry Thomson; a veto needs a two-thirds override. The 11-camera Feb 2024 Flock contract was already non-renewed and those units are removed. Still live on paper: 4 downtown security cameras and 4 mobile trailers ($90,000 Oct 2025 agreement); co-sponsor Isabel Piedmont-Smith expects the downtown four out within 60 days if Thomson signs, and says the trailers are in storage. IU Police and Indiana State Police ALPRs are outside city code. Monroe County already canceled Flock in July.
Why it matters
Split the SOW: city vs IU/ISP vs trailers-in-storage. Do not bid a Flock reboot; if Thomson signs, the next RFP is closed-circuit parks/body-worn, not networked ALPR.
Source
Indiana Public Media - 3
Key point
Downey City Council approved 4–0 (Councilman Hector Sosa absent) a five-year $532,026 Motorola Solutions ALPR expansion: 23 locations (13 existing + 10 new), paid in five installments of about $106,405. City owns the cameras at term-end and keeps paying Motorola software / plate-database access. Staff picked Motorola over Flock (18 cams $78.7k year-1 / $76k recurring, or 28 cams $220.2k over two years) and Axon (18 cams $269,910 / 5 years), citing 10+ years on Vigilant, investigator familiarity, and historical data without a second subscription. Ten older cameras are end-of-life; eight downtown 2023–24 cameras have one year left and can extend.
Why it matters
Overnight counter-example: a California city grew Motorola instead of killing Flock. If you are bidding Vigilant refresh, quote ownership-at-term plus the software tail, and do not assume a Flock-class privacy fight is on this agenda.
Source
The Downey Patriot - 4
Troy, NY: Mayor Mantello vetoes Local Law 3 (72-hour retain / court-or-DA share); council likely to override
Key point
Mayor Carmella Mantello vetoed Local Law 3, which the all-Democrat council passed 6–0 last month (already numbered 10-02). The law would force 72-hour ALPR deletion unless an officer has reasonable belief it is needed for a criminal investigation, and would allow sharing only under a court order or with the Rensselaer County DA. Mantello: the law is “biased in favor of criminal defendants” and “undermines” TPD; every chief and the DA testified against it. WAMC: the council will likely override. Separate fight: council is suing her over spring emergency-power Flock payments; she already lost a veto on emergency-power limits.
Why it matters
Hardware stays. The live question is whether 72-hour retain + DA-only share survives override. Do not bid a pull or a rewrite until the override vote lands.
Source
WAMC - 5
Providence unanimous referral of ALPR-ban ordinance to committee; 60 city Flock, $155k–$170k/yr; two more votes
Key point
Council unanimously referred the Miller / AnderBois / Sanchez ordinance to committee. This vote does not ban the cameras. It would prohibit operating ALPRs or using their data, require discontinue/terminate, and take effect on passage. PD runs 60 Flock; ordinance text cites 100+ in the city including RISP. Contract (Oct 2023) $155,000–$170,000/yr. Mayor Brett Smiley wants guardrails, not a kill; Chief Oscar Perez says he was not consulted. After committee + public hearing it needs two full-council votes, then mayoral signature or veto. Thursday’s 10-15 brief numbered the introduction only.
Why it matters
Treat as committee, not a dark. RISP’s cameras are a separate customer. Watch the next noticed vote before you price a pull or a swap.
Source
WJAR / TurnTo10 - 6
Key point
Orange Pi listed the O1, a 90 × 70 mm SBC on SOPHGO BM1688 (8× Cortex-A53 @ 1.65 GHz + RISC-V C906 + 16 TOPS INT8 TPU). VPU: decode 16 channels 1080p30 H.264/H.265, encode 10. Memory 8 or 16 GB LPDDR4/4X; M.2 NVMe, eMMC or SPI flash, SATA III, microSD; HDMI 2.0 4K, MIPI DSI, 2× 4-lane MIPI CSI; dual GbE, Wi-Fi 6E / BT 5.3, optional 4G/5G M.2 B-key; 2× USB 3 + 2× USB 2; 26-pin GPIO. Priced $195 (8 GB) / $325 (16 GB) plus $7.90 active cooler on AliExpress. CNX: the advertised Ubuntu image Google Drive is empty and docs are missing; claimed local Gemma-2B / Qwen1.5-1.8B / Llama2-7B / ChatGLM3-6B support is unverified.
Why it matters
Sixteen 1080p decode on a sub-$350 board is an NVR-class experiment, not a locked ALPR camera. Accelerated compute beats locked “AI cameras” (model swap vs 36–84 month refresh) — but do not quote it until the Ubuntu image actually boots.
Source
CNX Software - 7
Key point
MaLoW (Memory as Weights) stops stuffing compressed clips back into a streaming VLM’s context. A recurrent HyperNetwork folds incoming video into fixed-size LoRA updates so history lives in weights and the current frame stays as input. Trained on MaLoW-Instruct (381k QA pairs / 56k videos). On OVO-Bench backward split, +3.9 to +11.8 pp vs prior memory methods across backbones, modest latency. Also lifts Simpler WidowX VLA success +9.3 pp.
Why it matters
Live-VMS Q&A fails when history and the live tile fight for tokens. Weight-memory is the pattern to ask a Thor/Jetson vendor for — keep the last hour without dumping the archive into context.
Source
arXiv:2610.22070 - 8
Key point
ARC-SORT is tracking-by-detection with no training. An IMM Kalman covariance sets a per-track Expansion IoU search region (worse prediction → bigger match window). Close-call crossings get at most one corrective ID swap from post-crossing appearance (+1.0 HOTA on SportsMOT and DanceTrack, <1 ms/frame). Offline re-ID of players who left frame drops the motion gate and links by appearance only. 81.82 HOTA on SportsMOT test, fewest identity switches of 69 leaderboard entries as of Aug 2026, MOTA 97.09 (tied 3rd); DanceTrack 62.1 HOTA. ~50 fps on a laptop RTX 3060. Code promised.
Why it matters
Crowded-plaza ID switches are the operator ticket. A covariance-sized search window plus one post-crossing correction is a tracker wrap you can drop on ByteTrack/BoT-SORT, not a new detector RFP.
Source
arXiv:2610.22306 - 9
Key point
FALCON (WACV 2027) runs camera 3D occupancy on an AMD Kria KR260 (4.3 W idle, 6.6 W inference) and treats latency as a deadline, not a mean. PL compute is cycle-bounded; a watchdog publishes the previous output flagged stale if a frame is late. Proxy (6× ResNet-50 DPU / frame): 300k frames all delivered inside 100 ms + ≤10.2 ms. Full six-camera FlashOcc is still ~1.3 s/frame — not 10 Hz yet. Occ3D-nuScenes mIoU 29.88 (92.8% of an FP32 baseline the DPU cannot run). ResNet-50 100k-frame tail: KR260 1.5× median (2.0× on the six-inference proxy) vs Jetson Orin 4.4× and NVIDIA L4 23.6×. HLS, recipe, and traces released.
Why it matters
Pole occupancy that misses a frame is worse than a slower mean. If the RFP is “every frame before the deadline,” an FPGA watchdog beats a Jetson mean-FPS slide.
Source
arXiv:2610.22143