IVS Briefs
IVS / Edge AI Research Brief
Overall assessment
Signal level
High on a Permian-Basin unanimous terminate, an El Paso manager-lock, and a Texas Comptroller grant EO that names Flock, Motorola, and Axon in one sentence.
Notable trends
- Overnight LPR is Texas-heavy: Midland city-only pull (private/hospital stay), El Paso vendor-agnostic lock, Huffines grant-pot not DIR.
- Georgetown is suspend-and-review; Lebanon is the bagged-keep counter-vote.
- Edge this window is a Hailo-10 M.2 with 16 GB, not another Jetson blog recap.
- Wednesday cs.CV dump is 381 new; Lab is roadside occupancy + streaming VLM + tracker recovery + VAD evidence.
Gaps
No 48h VMS SKU (CyberLink FaceMe 8.9.2 hit did not resolve to a dated 8.9.2 release; Genetec/Milestone silent). No Qualcomm / DeepX module SKU (Hailo-10 is covered via Geniatech). No World Labs / NVIDIA Metropolis post in 48h (OCP Compal/Pegatron are datacenter racks). Golden study session and Downey Motorola $532k / 23 sites still unpublished. Ultralytics 8.4.177 (SAM prompts persist) left unnumbered as a patch on already-numbered 8.4.176.
Mix counts
LPR 3 / VMS 0 / Edge 1 / Policy 2 / Lab 4. LPR+Policy city/contract = 5/10 = 50%.
Now vs On deck
Now 6 (items 1–6) · On deck 4 (InfraWorld, SPARE, Dual-Agent, Temporal Evidence VAD).
Signals (10)
- 1
Midland, TX unanimous: terminate Flock; 73 city cameras, 60-day pull by Dec. 12, skip $225,500 year-2
Key point
Midland City Council voted unanimously Tuesday to terminate Flock. The city network is 73 cameras (71 fixed solar + 2 portable, inactive). 60-day written notice runs to Dec. 12. Ending now skips the $225,500 second-year subscription if notice lands before the Jan. 15, 2027 billing date; Flock refunds unused first-year fees pro-rata, with no early-termination fees, and removes/restores at its cost. Contract was extended February 2026 (two years). Mayor Lori Blong: this was the no-cost inflection before auto-renew. City IT/PD will preserve records required under TPIA / legal hold, then deactivate every city user. City cameras only — MRT lists 200+ community/private Flock remaining (Midland Memorial 8, Home Depot 5, Lowe’s 3, MISD, Midland County, DPS, HOAs, oil-field operators).
Why it matters
Quote the $225,500 year-2 cliff and Dec. 12 pull, then split the SOW: city ROW vs hospital / big-box / county / DPS that stay live. Do not bid a city-wide dark.
- 2
Key point
El Paso City Council unanimously passed the resolution that grew out of the Sept. 15 60-day removal (150 grant cameras). It bars the city manager from entering, renewing, or extending any agreement for fixed, networked ALPRs — Flock or “substantially similar” — unless council acts again. Handheld, vehicle-mounted, parking, access-control, and toll systems are carved out. Rep. Chris Canales: any return of the technology, any vendor, comes through council. The Sept. 15 order already put Flock on a Nov. 15-class removal clock, with cover-if-late. Chief Peter Pacillas has said losing the system will slow investigations.
Why it matters
This is a manager-lock, not a Flock-only ban. An Axon/Motorola/Verkada swap still needs a noticed council item. Bid only the carved-out mobile/parking classes unless council puts fixed ALPR back on the agenda.
Source
KFOX14; KTSM - 3
Texas Comptroller Huffines: executive order bars Comptroller grant money for Flock / Motorola / Axon ALPR
Key point
Comptroller Don Huffines signed an executive order banning Comptroller-administered grant funds to buy, subscribe, extend, or renew automated license-plate / vehicle-identification systems that convert plate images into computer-readable data. Named vendors include Flock Group Inc., Motorola Solutions Inc., Axon Enterprise Inc., and successors/resellers. The office will write the restriction into future solicitations and agreements. It is not a statewide ALPR ban and does not, by its terms, kill local general-fund buys or the DIR cooperative (DIR-CPO-5844 still listed active in public comment).
Why it matters
Texas RFPs that leaned on Comptroller public-safety grants now need a non-grant funding line. Ask which pot pays before you price Flock/Axon/Motorola; a DIR rider is not a Comptroller-grant waiver.
Source
Texas Comptroller - 4
Key point
Georgetown City Council voted 5–3 Monday to immediately suspend active Flock operations while it runs a full technology review. A narrower “police keep using, no outside share” option lost; the body took a full suspend. Councilmember Michael Crisp made the motion and said council will bring in independent legal, cybersecurity, and community experts plus public comment before a final keep/kill. Tammy Mitchell and Karen Sames wanted a serious conversation before any contract move; Greg Hampton called the vote an overreaction.
Why it matters
Treat Georgetown as bagged pending expert review, not a terminate. The next agenda item is the review packet, not a swap RFP.
Source
LEX18 - 5
Key point
Councilmember Matt Sellers’s terminate resolution failed 4–3 Tuesday. Lebanon’s 20 Flock cameras stay in place and stay covered — they were shut off and bagged last month after Flock missed a 50-question deadline (a later reply still left Mayor Mark Messer uneasy on retain and security). After the fail, all seven members backed Kristen Eggers’s motion directing staff to coordinate ALPR-policy talks with public-safety officials. Mike Cope (no on terminate) wants county alignment and contract language on deletion, not a reboot of the current terms.
Why it matters
This is the overnight counter-vote: hardware stays, video stays off. Do not bid a pull or a turn-on until staff comes back with a rewritten data-deletion rider.
Source
WVXU - 6
Geniatech M.2 cards: AIM-M-R28 (RK1828 20 TOPS, 5 GB stacked DRAM) and AIM-M-H10 (Hailo-10 40 TOPS, up to 16 GB)
Key point
Geniatech listed two M.2 2280 PCIe accelerators for on-card LLM/VLM/CV. AIM-M-R28: Rockchip RK1828, 20 TOPS INT8, 5 GB 3D-stacked in-package DRAM, PCIe 2.1 x1, 20 W TDP, −25 to 85 °C, hosts RK3568 / RK3576 / RK3588, Linux/Android, claimed 3B–7B local LLM. AIM-M-H10: Hailo-10 40 TOPS, up to 16 GB LPDDR4/4X, 12 W typical, 0–70 °C, Linux/Windows, hosts NXP / NVIDIA / Qualcomm / AMD Xilinx; Hailo Dataflow Compiler + Model Zoo (CNN, Transformer, ViT). Versus ASUS UGen300 USB Hailo-10H (8 GB), the H10 is internal M.2 with more RAM. Public price/availability not posted — sales-inquiry SKUs.
Why it matters
This is a swappable M.2 NPU, not a locked ALPR camera. Accelerated compute beats locked “AI cameras” (model swap vs 36–84 month refresh).
Source
LinuxGizmos - 7
Key point
InfraWorld (roadside occupancy forecasting) keeps a dense BEV state for the whole scene plus persistent motion queries on moving regions. DQI seeds queries where history showed motion; LST fuses short-frame continuity with long-term trends; MAQI injects confidence-weighted query updates into BEV tokens without wrecking global context. SOTA on InfraOcc and Occ3D-nus; still works in ego-centric settings. Code promised.
Why it matters
Pole-mount / intersection occupancy is the infrastructure-camera job, not a dashcam paper. Watch the InfraOcc split before you promise 4D “what moves next” on city poles.
Source
arXiv:2610.14544 - 8
Key point
SPARE (Shared Perception and Adaptive Retrieval of Evidence) answers streaming-video questions from recent frames, pulling older clips only when a lightweight gate says recent evidence is not enough. Understanding and retrieval share a visual encoder (separate language backbones; four linear adapters distilled at token and embedding level). LoRA trains the model to use both streams and abstain when evidence is still thin. Visual-only: 77.78% on OVO-Bench Real-time/Backward groups and 73.31% on StreamingBench — +20.68 / +14.15 pp vs Qwen3-VL-8B-Instruct at 2 fps.
Why it matters
Live-VMS Q&A fails when the model either forgets the last hour or dumps the whole archive into context. The gate-and-retrieve pattern is what you want on a Thor/Jetson stream, not a 2 fps cloud VLM.
Source
arXiv:2610.14671 - 9
Key point
SAM-family VOS trackers propagate masks through memory; one bad frame under occlusion/reappearance/distractors locks in identity drift. Dual-Agent adds a VLM verification agent (is this still the target?) and a correction agent trained on real SAM3 failures plus identity-switch/distractor perturbations. +1.2 / +2.7 J&F on LVOS-v2 and SecVOS vs SAM3; transfers without retraining to SAM2.1, SeC, Cutie, and long-term MOT (LaSOT-ext, TNL2K, VOT-LT2022).
Why it matters
ReID/MOT on poles dies on the first occluded bus. A verify-then-recover loop is the ops pattern — wrap the tracker you already licensed rather than rip it out.
Source
arXiv:2610.15929 - 10
Temporal Evidence Selection: VAD explanations need object-tag–conditioned sampling, not anomaly saliency
Key point
Video anomaly understanding (language-level explanations of long videos) usually samples the “suspicious” frames and misses context, interactions, and distributed evidence. A semantic-conditioned Temporal Sampling Q-Former learns what to keep from the generation loss, gated by CLIP-embedded object tags taken from the video (tags condition the sampler only; they are not stuffed into the LLM prompt). Soft temporal weights in visual-token space; duplicate selections merged. Best among compared fine-tunes on HIVAU-70K; ablations beat fixed anomaly-focused sampling.
Why it matters
“Show me why that’s an anomaly” is the operator ticket, not a heatmap. If your VAD narrative only replays the peak-score second, this paper is why the explanation fails in review.
Source
arXiv:2610.16299