IVS Briefs
IVS / Edge AI Research Brief
Overall assessment
Signal level
High on the Houston split (Manvel kill / Conroe keep / Willis pause) and on Flock’s AI-training answer. High on open VMS (Milestone 5,800).
Notable trends
- Same metro, three products: terminate, budget-keep, pause.
- Vendor admitted customer video trains the model.
- WI FOIA of 2025 case files is now a sheriff’s “lied to” letter.
- GSX week is open VMS + camera OEM mergers, not another ALPR pole.
Gaps
No new Hailo / DeepX / Qualcomm / Jetson SoC. Chroma24 is an imager, not an NPU. Collegedale primary was thin. day-range was noisy (27 unique URLs on 4 queries); Lab pass + GSX PRs carried Edge/VMS.
Mix counts
LPR 3 / VMS 2 / Edge 1 / Policy 0 / Lab 2 (LPR+Policy = 3/8 = 38%).
Now vs On deck
Now 5 (items 1–4, 6) · On deck 3 (March GSX, MMGait, Point4D).
Signals (8)
- 1
Key point
Manvel City Council 6–0 to terminate. Staff: 30-day cancel notice; cover 16 operating cameras in 24–48 hours (2024 contract was 21). Mayor Dan Davis asked Flock’s Trevor Chandler (Zoom) if Manvel data trains Flock AI. Chandler: “part of making our cameras more accurate.” Davis: “That’s a yes.” Chandler: no data sale, no FR. PD: 535 outside agencies still had standing access after the city had already cut 149 (mostly TX, rest LA). Chief Keith Traylor: last monthly audit 138 local searches, no intentional misuse; could not say how many closed a case. Ending the contract also cuts Manvel’s inbound search of other agencies. Davis cited Sheboygan Traffic Analytics and Pflugerville’s 459-agency / 1.6M-search miss.
Why it matters
New RFP line: does customer video train the vendor model? If yes, that is a records and IP clause, not a privacy slogan. 535 remaining partners after a “cut” is the sharing-graph job.
Source
ABC13; Texas Scorecard - 2
Key point
Sheriff Travis Maze voluntarily suspended Flock. When the county expanded in 2025, Flock told him data would not be released without permission. This week he learned 2025 investigations were made public without consent (same FOIA-of-search-logs pattern as Fond du Lac / Winnebago). No local officer-misuse finding. “This suspension is based purely on them telling us one thing about how our data would be accessed, and then having the exact opposite happen.” Awaiting Flock’s explanation. Atlas of Surveillance: ~10 Jefferson County SO cameras.
Why it matters
The WI packet is now MSA lie + FOIA of active cases, not just Traffic Analytics. Integrators: name the records custodian and what happens when another state’s public-records law hits the audit log.
- 3
Key point
Same Houston week as Manvel’s 6–0 kill: Conroe left Flock in the operating budget. No separate Flock vote — funding survived inside the $389.9M FY26–27 budget (3–2: Blevins / Hoisington / Sellars yes; Harrell / Arthur no). Chief Jon Buckholtz: 51 cameras; ~894,000 unique scans; ~642 searches. “That’s not mass surveillance.” Harrell: contract has a non-appropriation clause — stop paying this budget year with written notice, no penalty; he wanted a pause. He and Arthur voted for Flock last year “without all the facts.” Mayor Duke Coon put support for the chief on the record. Willis (north of Houston) shut cameras until further notice Sept 8 — review, not terminate.
Why it matters
The keep-slide is now search/scan ratio (642 / 894k) plus a no-penalty non-appropriation clause most cities never use. Bid both forks in the same metro: Manvel is dark, Conroe is 51 live.
Source
KPRC - 4
Key point
Open-platform XProtect is the spine for an 88-acre Strip property (3,500 rooms, 117k sq ft casino). Mix: Axis, Bosch, Hanwha cameras; BriefCam forensic / people-count / heat map; Oosto FR; Aeyesky card-count/cheat detect; AXIS Case Insight foot traffic. Assistant director of fraud/nightclub surveillance: daily use for fraud, theft, loss. Milestone Americas VP: open platform so best-of-breed does not add operator complexity. Next: more AI people-count, behavior, guest-path tools.
Why it matters
This is the casino/hospitality RFP: open VMS + specialist analytics, not a locked AI camera pack. If you sell XProtect, the attach list is BriefCam/Oosto/Case Insight — quote the stack, not a dome.
Source
Security Info Watch - 5
Key point
Five months after the April merger (both under Delta Intelligent Building Technologies), first combined GSX. Demos: broader VIVOTEK cameras in March environments; C•CURE events inside March Command Enterprise; VORTEX cloud into Searchlight Cloud; Searchlight AI natural-language search included (no extra license) for Searchlight Cloud customers; expanded face + LPR search. Named win: 500+ store North American apparel retailer — VIVOTEK cameras + VORTEX + March ops intel vs a competitor cloud-analytics bid.
Why it matters
Retail/finance March shops just got a camera OEM inside the family. NL search with no SKU tax is the close against Genetec/Milestone cloud packs. Walk booth 1823 if you already sell either brand.
Source
PR Newswire - 6
Key point
Same GSX drop: Chroma24 is the imaging SKU — full-color video in ultra-low light so AI detection and video search still fire at night. Positioned as the sensor that makes Searchlight/VORTEX analytics honest after dark, not a new NPU. On the truck this quarter via VIVOTEK/March channel.
Why it matters
Night color is the difference between a usable LPR/person search and a grayscale miss. Prefer this class of imager on a swappable analytics box over a locked “AI camera.” Accelerated compute beats locked “AI cameras”: swap models vs a 36–84 month hardware refresh.
- 7
Key point
Gait ReID is usually RGB/silhouette. MMGait aligns visible, IR, depth, LiDAR, and radar at sequence level. OmniGait++: modality-specific fronts + shared ID encoder; fusion that works when a sensor drops out. Cross-modal retrieval is still hard; fusion helps; one checkpoint covers single / cross / multi-modal instead of a zoo of experts.
Why it matters
Perimeter sites already mix thermal + RGB + radar. A ReID pitch that dies when the IR camera is the only working view is not a product. Ask vendors what happens when one modality is missing.
- 8
Key point
Current 4D reconstructors choke after a few dozen frames. Point4D predicts dense per-point 3D trajectories on multi-hundred-frame video via a 3D query decoder that does not require the point to stay visible in the image plane; endpoints re-query the next chunk with no re-projection/matching. Reusing a descriptor from any visible frame beats the source patch. SOTA on long-video tracking >200 frames. Project: point-4d.github.io.
Why it matters
Occupancy, lofted tracks, and “where did that cart go for the last three minutes” are long 3D jobs. Spec a model you can swap on Jetson/Thor, not a camera firmware that only tracks inside a 2-second clip.