IVS Briefs
IVS / Edge AI Research Brief
Signals (8)
- 1
Lockport 5–1 revokes Flock permission; 2 cameras authorized in February; sponsor now rejects Axon swap
Key point
Lockport Common Council voted 5–1 to revoke prior permission for all Flock ALPRs in the city. Fourth Ward Alderman Kathryn Fogle sponsored the resolution; Council President Rowland O’Malley (3rd Ward) cast the lone no. The city had authorized two cameras in February (Summit & Transit; East Avenue & Davison Road). Fogle said the Niagara County Sheriff’s Office will need to remove them. After the vote she said she no longer wants Flock replaced by another brand, “not even by Axon now,” even though the city already runs Axon readers in police cars. A sheriff’s investigator warned that crime does not stop at municipal lines and that the regional system costs the city nothing. WHEC: the resolution also cancels planned installs.
Why it matters
Western New York now has a 5–1 revoke with no-swap language. Bid a local VMS against two sheriff-owned poles and an in-car Axon stack the sponsor just took off the table; do not treat the vote as hardware already down.
- 2
Dutchess misses Oct 30 Flock remove: 35 / 64 down or in process, 22 need DOT permits, 7 never installed
Key point
After the Sept. 30 demand notice set an Oct 30 hardware deadline, Communications Director Colleen Pillus said Friday that 35 of 64 county Flock cameras had been removed or were “in the process of being removed,” 22 need New York State DOT permits, and 7 were never installed. No date for the rest. DA Anthony Parisi’s office and Sheriff Kirk Imperati’s office each pointed elsewhere; Flock’s media line went to voicemail. Cameras went offline Sept. 25. The sheriff’s Flock portal still shows 30+ searches since then; Capt. John Watterson said those hits were on privately or municipally owned cameras, not the county set. Legislature cover resolution takes effect Nov 4. Do not claim the remaining 22 are covered.
Why it matters
A county go-dark that still has poles up is the Hudson Valley unwind pattern. Quote a 22-camera DOT-permit remaining set plus a Nov 4 cover order; city/town cameras (Poughkeepsie, Beacon) are a separate bid.
Source
Times Union - 3
Dayton, MN: Flock reinstalled two cameras after cancel; 60 days to pick up; Grass Valley work-order pattern
Key point
Dayton, Minnesota canceled Flock in August. City Facebook: PD saw a Flock maintenance work order Sept 17; the city told Flock Sept 18 the program was on hold and no work orders were to be completed. Flock put two cameras back into service anyway. Residents spotted them; police were directed to remove them. Administrator Zach Doud: Flock took five days to acknowledge the cancel and said retrieval will take 60 days. KARE 11: St. Paul and Crystal still hold cameras in storage; Crystal Chief Pete Underthun removed poles himself. TechSpot: Grass Valley, CA saw the same reinstall after an Aug 11 cancel letter — Flock said city self-removal looked like equipment failure and auto-generated replacement work orders.
Why it matters
Unwind RFPs now need a written no-work-order / no-replace clause and a city-held hardware pickup, not a vendor SLA that can dispatch crews after a vote. Quote Dayton’s 60-day retrieve against Crystal’s self-remove.
- 4
Key point
League City voters decide Proposition A on Nov 3: whether PD should continue using 55 Flock ALPRs on public roads in a city of about 120,000. It is a local advisory question (Ballotpedia: LRAQ), not a statute. Flock has been defending its safeguards ahead of the vote. ACLU of Texas senior manager Nick Hudson: those protections fall short for drivers who are not suspected of crimes. York, Maine has a separate Flock-removal initiative on the same day. Do not claim a result.
Why it matters
First voter-facing ALPR question of this cycle is a Houston-suburb 55-camera advisory. Have a local retain/VMS price ready for Wednesday; an advisory yes does not lock a contract, and an advisory no still needs a council PO.
- 5
BrainChip AKD1500 Raspberry Pi 5 Dev Kit: neuromorphic YOLO/person/face demos, $1,295, 800 GOPS / <300 mW
Key point
BrainChip shipped a boxed AKD1500 Raspberry Pi 5 Dev Kit: Pi 5 8 GB, official M.2 HAT+, Akida AKD1500 M.2 B+M-Key card, metal case, 32 GB SD with Ubuntu Server, Python, and MetaTF. USB camera in; browser launcher for Visual Wake-Word person presence, YOLO object detection, MobileNet edge, and MobileNet face demos — all on-device. AKD1500: Akida Neuron Fabric 5–400 MHz, 1 MB on-chip memory, 800 GOPS at <300 mW, on-chip class learning, no cloud. Kit $1,295. CNX notes the same parts à la carte are about $350 (Pi 5 8 GB $175, HAT+ $12, M.2 card $129). M.2 card was already $129 in September; this is the first complete vision demo kit.
Why it matters
A sub-watt event-based NPU with a YOLO/person/face demo is a low-power presence node you can swap models on, not a locked AI camera. Accelerated compute beats locked AI cameras (model swap vs 36–84 mo refresh).
Source
CNX Software - 6
Key point
Zeping Zhang, Zhuoya Zhao, and Robert Laganière show randomly dropping 75% of vision tokens already beats a dense driving model. GATS (Geometry-Aware Token Selection) uses LiDAR and radar point projections to keep a small fraction of image patches before the backbone runs, then anchors each kept token in 3D for a self-supervised latent world model (no box labels). At 25% budget: nuScenes 3-sec L2 0.520 m and NAVSIM EPDMS 88.9 vs dense 0.573 m / 85.9 at 6.6× fewer FLOPs. At 10% budget (24 GFLOPs, 18× cut) it still beats dense. On 739 intent-critical scenes it covers 95.8% of hazardous vehicles with a 0.09% collision rate vs 0.34% dense.
Why it matters
Perimeter stacks that already have a LiDAR can starve the GPU of sky and empty-road tokens. Quote GATS-style geometry gating when an occupancy RFP asks for camera+LiDAR fusion without a second training campaign.
Source
arXiv:2610.40935 - 7
Key point
Same authors ship UFMF (Ultra-Fast Multi-modal Forecaster): an attention-free diagonal state-space core plus a radar-guided velocity-aware dynamic prior that splits future displacement into linear extrapolation plus a learned residual, then transports a compact dynamic latent during rollout. 122 FPS on one RTX 4090; 0.45 fast-displacement mIoU on nuScenes, beating heavier occupancy world models at a fraction of the compute. World-frame dynamic supervision and class-conditioned vertical-profile loss regularize BEV-to-3D lifting.
Why it matters
Occupancy that holds fast movers at real-time rates is the perimeter forecast layer to quote against generative 4D world models that miss a truck latency budget.
Source
arXiv:2610.40947 - 8
TwinPast: streaming VLMs that pass recent-frame quizzes fail when history actually changes (49.14% PairAcc)
Key point
Haotian Yang et al. find 32% of sampled “historical” questions on OVO-Bench / StreamingBench are answerable from recent frames, and models also score from the question text alone. TwinPast holds recent video, question, and option order fixed while swapping the historical evidence and the correct answer across 350 human-reviewed pairs / five tasks. PairAcc requires both histories correct. Strongest streaming system: 49.14% PairAcc vs 66.57–91.43% for general-purpose models. Models pick the wrong time or contradict event order even when the evidence is in memory.
Why it matters
SOC copilots that look accurate on a live stream can ignore the clip from 10 minutes ago. TwinPast PairAcc is the evidence-readiness number to put in an RFP when a VLM is asked who entered, left, or returned.
Source
arXiv:2610.41983