IVS Briefs
IVS / Edge AI Research Brief
Signals (8)
- 1
Riverside, IL 4–2 turns off 2 Flock; 3 Verkada $19,779 by Jan 1 2027; Flock software stays for ~30 cases
Key point
Village board 4–2 (Evans, Fitzgerald, Mateo, Marsh-Ozga yes; Gallegos, Kos no) turned off the two Flock units at Longcommon/Byrd and Burlington west of Harlem on Thursday. Earlier the same night, 4–2 authorized three Verkada remote mobile cameras via Griffon Systems Inc.: $6,593 each (hardware, power, install, one-year license), $19,779 total, delivery/install by Jan 1 2027 if on plan. Two replace the Flock poles; a third is new at Des Plaines/31st. A later 4–2 directs PD to terminate Flock data-sharing software only when the Verkada units go live. Public Safety Director Matthew Buckley and Cmdr. James Lazansky: cutting Flock now would dump ~30 live investigations. Marsh-Ozga opposed the third camera.
Why it matters
A two-cam Flock dark / three-cam Verkada swap with Flock software held for cases. Quote the Jan 1 2027 cutover, not Thursday, and price a local VMS/LPR against $6,593/cam plus the third-site add.
Source
RB Landmark - 2
Milford, CT alders unanimous: expansions and renewals of surveillance tech need the Board and public comment
Key point
Board of Alders unanimously passed a resolution: any expansion or use of surveillance technology, and any contract renewal, must come to the Board with a public-comment window; 30 days’ notice for other new tech. Police chief must list every system in use, cost, and vendor within 90 days. City had paused nearly all cameras in August after Flock/Condor backlash; Condor paused Oct 5. Democrats and Republicans backed oversight even while they still split on whether Flock should come down. This is an oversight lock, not a ban.
Why it matters
Do not bid a Flock/Condor/drone add into Milford without a Board hearing. The 90-day inventory is the RFP map; the next fight is still removal vs keep-with-rules.
Source
NBC Connecticut - 3
Excelsior Springs: 15 bagged Flock still capturing plates; second bags; Flock-owned, contract ends this month
Key point
After the Oct 19 unanimous terminate (already 10-20), public works bagged the city’s 15 Flock cameras. Fulfilling a Sunshine Law request, Lt. Ryan Dowdy found the units were still capturing as recently as Wednesday. Most frames are headlights through nylon; some plates and vehicles are readable. Crews are adding a second bag. City does not own the hardware; Flock is expected to pull it when the contract ends this month. PD has not used post-vote images. Mayor Sonya Morgan had ordered bags as the immediate step; Chief Gregory Dull said detectives go back to standard cameras and tips.
Why it matters
A bag is not a disable. Write take-down quotes that include Flock-owned pickup plus a remote decommission check, and do not certify “dark” from a photo of a trash bag.
Source
KCTV - 4
Quincy swaps Flock for OdinEye at $250–$500/cam; vendor is a Middletown, CT detective (CellTrace LLC)
Key point
Four days after the recommend-remove (already 10-20), Mayor Thomas Koch (Oct 23) hired OdinEye with no public bid. Software-only: it aggregates plates from third-party cameras so agencies can search across vendors. ~$250–$500 to outfit a camera vs Flock’s ~$3,400 hardware+software+maintenance bundle. Promises E2E encryption, case-number gates, audits. Globe: the only identified principal is John Stroud, a Middletown, CT detective, doing business as CellTrace LLC (2024, sole employee). LinkedIn: 850+ users across 40 Connecticut agencies. Councilor Ziqiang Yuan and another member filed a late-October ordinance to require council approval of surveillance tech. ACLU of Massachusetts wants a public hearing first.
Why it matters
The Flock hole is being filled by a software aggregator, not another pole SKU. Bid camera-agnostic LPR search against $250–$500/cam, and diligence CellTrace as a one-person shop before a citywide cutover.
Source
Boston Globe - 5
Key point
Mouser is stocking Advantech’s MIC-743-AT inference box on NVIDIA Jetson Thor T4000/T5000. Up to 2,070 TFLOPS (FP4) for LLMs, VLMs, and video analytics; 14-core Arm Neoverse V3AE, 128 GB LPDDR5X, 1 TB SSD, 2,560 CUDA / 96 5th-gen Tensor cores at 1.57 GHz MAXN. HDMI, USB 3.2 Gen2, M.2 A+E (Wi-Fi), M.2 B (LTE). JetPack 7.0 preinstalled. 19–36 V in. Operating −10°C to 35°C with 0.7 m/s airflow. Companion “development boards” listed at 1,200 TFLOPS FP4. No street price in the Mouser write-up. Indoor/closet thermal envelope, not a −20 to +70 Pico-ITX.
Why it matters
A Thor-class VLM/video box you can order through Mouser for a rack or closet NVR, with model-swap headroom the locked ALPR camera does not have. Quote it only where you can cool −10 to 35°C with airflow; accelerated compute beats locked “AI cameras” (swap models vs a 36–84 month refresh).
Source
Engineering.com - 6
Key point
Phuong T. Nguyen et al. formalize Refer-MTMCT: a free-form query must recover identity-consistent tracks across a sparse-overlap camera network. Refer-CityFlow (on CityFlow): 96 videos, 40 roadside cameras, 660+ cross-camera trajectories, 4,362 descriptions, 7,699 queries including 1,699 real-user queries from 50 people. On four-camera Scenario 02, single-camera referring trackers hit 14.95–18.28 HOTA on real-user queries; a cross-view baseline in direct transfer is ~0.05 HOTA. M3Track, an LLM agent over interchangeable detect/track/associate modules, reaches 44.10 HOTA and 65.40 IDF1. GitHub posted.
Why it matters
“Find the red pickup that left gate 3” is now a benchmark, and naive cross-camera transfer is near zero. Ask MTMC vendors for language-query HOTA on a real user set, not only IDF1 on scripted IDs.
Source
arXiv:2611.08976 - 7
Key point
Sadique Khatib et al. compute a frame-level Thermal Reliability Coefficient (TRC) from five unlabeled cues and use it to gate thermal features in a dual-branch YOLOv8m (visible branch frozen, COCO-pretrained) plus ByteTrack. M3FD test: 87.6% precision, 79.4% recall, 70.0% mAP@50, 47.8% mAP@50:95; 17.0M / 42.9M params trained. TRC ~0.78 on clean frames, 0.53 / 0.69 under full thermal saturation/blur. Adaptive gate matches a fixed gate on clean data (70.0 vs 69.8) but holds 43.6% mAP@50 when thermal is corrupted vs 38.4 / 29.2 / 25.0 for fixed / full-thermal / closed-gate (+5.1, 95% CI +3.6 to +6.5). Tracking gain is not shown; benefit is the detector.
Why it matters
Dual-imager RFPs fail when the thermal path saturates at noon. Ask for a reliability gate with a corrupted-thermal mAP, not a single clean-set number.
Source
arXiv:2611.09725 - 8
Key point
Dongxiao Li, Zixuan Hu, Ling-Yu Duan argue spatial-domain TTA on sparse LiDAR confuses semantic content with sensor/style shift. FIND (Frequency INvariance Disentanglement) fits B-splines into locally adaptive filters: an Invariant stream drives pseudo-labels, a Specific stream regularizes against domain perturbations. Dual-stream, no retraining at test. Outperforms prior TTA on cross-dataset and corruption benchmarks. Code: github.com/llldx03/FIND.
Why it matters
Roadside LiDAR RFPs change sensor and weather between demo and go-live. Ask whether the detector adapts in frequency at test time, or was only trained on the integrator’s one OS1.
Source
arXiv:2611.09170