IVS Briefs
IVS / Edge AI Research Brief
Overall assessment
Signal level
High on an overnight 7-day / no-video first vote (West Lafayette), a 7–0 remaining-camera kill (Village of Milford, MI), a school-police Flock MOU refusal (CCSD), and Raspberry Pi’s official 25 TOPS / 3 W DEEPX HAT post.
Notable trends
- Overnight ALPR action is retain-and-scope (West Lafayette) as often as non-renew (Milford MI).
- School police can drop Flock without owning a pole.
- A 3 W NPU HAT is now an official Pi News SKU, not a June DeepX recirculation.
Gaps
No 48h VMS SKU (Columbia Milestone reprint already numbered 09-28). No new Hailo / Qualcomm module (Pi/DEEPX HAT fills Edge). Tewksbury terminate left un-renumbered (10-01). Scottsdale / Fountain Hills / Apache Junction / East Peoria / Tewksbury bylaw are tonight — not padded as a seventh Flock pullout.
Mix counts
LPR 2 / VMS 0 / Edge 1 / Policy 2 / Lab 3. LPR+Policy city/contract = 4/8 = 50%.
Now vs On deck
Now 5 (items 1–5) · On deck 3 (TRIM-ReID, RBA VAD, referring MOT).
Signals (8)
- 1
Key point
West Lafayette City Council unanimously passed first reading of tighter Flock rules. Data storage drops from the 30-day window in 2022 city regs to seven days unless it is tied to an ongoing investigation, and the ordinance bars expansion into video recording, live feeds, or voice files. Four Flock cameras stay in place. Residents pressed for a full ban by the Nov. 2 second/final vote. Council member David Sanders, one of the two sponsors, said the first-reading package is small and he is open to ending the cameras.
Why it matters
This is the overnight vote Monday’s brief was watching. Quote 7-day delete + no-video/no-live/no-voice as the floor, and keep a Nov. 2 teardown bid ready if Sanders and the gallery convert the second reading into a pull.
Source
Based in Lafayette - 2
Village of Milford, MI 7–0: remaining 10 Flock ALPRs will come down when the contract dies this month
Key point
Village Council voted unanimously not to renew Flock. The village approved 10 cameras in 2023 and five more in 2024, then killed the five-camera contract in June; Monday’s vote removes the original 10. The contract expires later this month, then the units come down. Police Chief Scott Tarasiewicz said the cameras helped on a homicide, break-ins, and missing persons, but downtime, vandalism, and poor vendor response drove his recommendation to exit. Residents had weekly protests and a 600+ signature petition. St. Clair Shores and Lincoln Park already left.
Why it matters
Metro Detroit is now a replacement cluster. Bid a cover-and-return plus a written deprovision date that matches the late-October lapse, and do not assume “cameras worked” beats a support SLA the chief will say in public.
Source
ClickOnDetroit / WDIV - 3
Key point
CCSDPD Chief Jeff Clark (confirmed July) said after a review that started with an August pause that the department will not pursue a new Memorandum of Understanding to access Flock. CCSDPD does not own or operate Flock hardware and had used partner-agency access since August 2024; a district memo said there are no Flock cameras on school property. Future updates go to the Board of School Trustees in public. Clark reserved partner tools for imminent-danger emergencies.
Why it matters
School-police ALPR is often a data MOU, not a pole count. Quote no-network-login / no-partner-share in K-12 public-safety RFPs, and do not sell Flock access as a camera-free add-on to a district that just walked it back on the record.
- 4
Key point
Atkinson commissioners unanimously prohibited installing or operating ALPRs inside town limits, citing N.C.G.S. §160A-174 and Fourth Amendment concerns. The ban covers any vendor, including Flock. Atkinson has no police department and no Flock poles in town; Pender County Sheriff runs eight Flock cameras, all outside Atkinson. New Hanover County has terminated Flock in parts, but no other NC city had banned the cameras. Population <300, one square mile. DeFlockILM’s Mark Williams spoke at the Thursday meeting.
Why it matters
A no-PD ordinance ban is a different product from a contract terminate. County sheriffs can still ring a town that just wrote ALPRs out of the municipal code — price county-only coverage and do not assume a 300-person ban is a camera RFP.
Source
WECT - 5
Key point
Official Raspberry Pi News post (Ashley) on the third-party Sixfab AI HAT+ for Pi 5: DEEPX DX-M1M NPU, 25 TOPS at ~3 W typical sustained NPU power, 2 GB on-module LPDDR4x. One NPU for CNN (YOLO ModelZoo + DX-Stream GStreamer), compact VLM, and SLM. Measured on Pi 5 8 GB: MobileNetV2 240×240 2361 FPS; DeepLabv3+ 512×512 155 FPS; Qwen3-1.7B TTFT 599 ms / 4.64 tok/s. Stack: DX-COM → .dxnn → DX-RT C++/Python. 25 TOPS HAT shipping now from Sixfab; 13 TOPS version expected late 2026.
Why it matters
A 3 W NPU HAT is a smart-camera / occupancy box you can actually thermal. Accelerated compute beats locked “AI cameras”: swap models vs a 36–84 month hardware refresh.
- 6
Key point
Wanke Xia et al.: multi-modal object ReID across RGB, NIR, and TIR. Visual encoders trained for global image-text alignment drop local identity and keep duplicate tokens. TRIM-ReID stacks Dense Identity Representation (DINOv3 patches), Token Diversity Mining (drop repetitive patches, keep complementary locals), and Modal Relational Interaction with a triangular alignment loss. SOTA on RGBNT201, RGBNT100, and MSVR310. Under review.
Why it matters
Night / thermal ReID is the after-dark search customers ask for once RGB hits glare. This is the token-budget recipe, not a shipping SKU.
Source
arXiv:2610.04361 - 7
Key point
Chao Huang, Pengfei Wei, et al. (Sun Yat-sen / University of Macau): MLLM VAD often knows normal vs abnormal in hidden states but says the wrong thing (representation–behavior misalignment). Direction of the native readout axis dominates the gap. RBA adapts the mid-to-late attention layers that actually separate the classes, using video-level labels only and updating ~0.012% of backbone parameters. One generative path for the anomaly decision and the explanation. Three VAD benches.
Why it matters
This is the “why did it flag” layer operators want on VAD, without a second classifier. Watch it as an adapter on whatever VLM you already run on the box — not a new camera.
Source
arXiv:2610.05129 - 8
Key point
Hsin-Chen Pai, Jyun-Kai Wang, Yi-Cheng Peng, Wei-Ta Chu: referring multi-object tracking takes a video plus a language query (“the vehicle that is turning”). On a moving camera a parked car looks like it moved. They estimate camera motion in driver-view video, extract residual object motion, match it to the query, and late-fuse with a host RMOT predictor. The compensation module is a plug-in across RMOT hosts.
Why it matters
PTZ, body-worn, and dash are moving cameras. If the query is about how something moved, you have to subtract ego-motion or you track the background.
Source
arXiv:2610.04185