IVS Briefs
IVS / Edge AI Research Brief
Signals (9)
- 1
Key point
Sherman City Council voted 7–0 on Resolution 7682 not to renew or extend Flock Group Inc. Mayor Shawn Teamann: the vote was a direct result of public outcry on privacy and constitutional issues. Next step is to alert Acting City Manager Clint Philpot to remove the city’s 11 ALPRs; Teamann said the measure is effective immediately, removal “takes a little bit of time.” Any future city-department ALPR install needs a public hearing. Councilman Clay Barnett voted yes while saying he hated taking a tool from PD; he cited data once it leaves Sherman. The resolution covers city-owned cameras only — Teamann said state and county ALPRs inside city limits stay.
Why it matters
Texas no-renew plus a hearing lock after the Comptroller grant ban. Do not bid Flock or a quiet Axon swap in Sherman — any plate stack comes back through a noticed hearing. County and state readers on the same roads are still live.
Source
KXII - 2
Excelsior Springs, MO unanimous: terminate Flock; 15 cameras bagged this morning; manager has 30 days
Key point
Council unanimously approved an ordinance that stops FY27 money on the Flock master services agreement and directs the city manager to terminate within 30 days. Mayor Sonya Morgan: public works bags the 15 cameras Tuesday morning; the agreement ends in November and Flock pulls the hardware. Morgan said a tight 2027 budget made the cut easy once residents spoke. Organizer Tracie Hardesty (Excelsior Springs Citizens Against Flock; 1,000+ petition signatures) thanked council. Police Chief Gregory Dull opposed the pull, will comply, and said the city policed without ALPRs before. KSHB: Clay/Platte vandalism and a Sept. 15 town hall preceded the vote.
Why it matters
Hardware is dark today, not at a future expire date. Bid a local-only retain/share stack only if council wants a replacement; Morgan’s line is budget-plus-privacy, not a vendor swap.
Source
KCTV; KMBC - 3
Key point
After more than an hour of public comment, council voted on a resolution recommending the mayor’s administration remove or cover the city’s Flock cameras. Order 2026-143 (Councillors Ziqiang Susan Yuan and Virginia M. Ryan) calls for terminating all Flock contracts, covering until removal, and council approval of any future ALPR. PD runs 49 ALPRs plus 1 Condor dump camera (Chief Mark Kennedy); Flock since 2020. Boston 25 did not publish a roll-call. The resolution does not itself pull the cameras — hardware stays until the administration acts. Companion Ordinance 2026-142 (public hearing before new surveillance buys) had no reported result.
Why it matters
Forty-nine live ALPRs is a South Shore-scale unwind if the mayor complies. Do not quote a swap until the administration’s written response.
- 4
St. Albans, WV: Council pauses any Flock / ALPR purchase and asks the Legislature for statewide rules
Key point
St. Albans City Council adopted a resolution pausing the city’s ability to purchase Flock cameras or any other ALPRs. Vice Mayor Jared Page: the city will not buy “any products of the kind” until statewide safety and privacy guardrails exist; PD was told and cooperated. The resolution also urges state lawmakers to set those standards. WSAZ did not publish a roll-call. Next council meeting is Nov. 2. This is a purchase freeze, not a terminate of cameras already in the field.
Why it matters
West Virginia is now a do-not-ship city until the Legislature moves. Do not bid Flock, Axon, or Motorola plate kits into St. Albans on a PO; any ALPR comes back after state rules, and Nov. 2 is the next local date.
Source
WSAZ - 5
Ultralytics 8.4.183: Axelera Voyager SDK 1.9.0 exports every YOLO26 task; Intel LiteRT CPU/NPU; 109 ips @ batch 32
Key point
Ultralytics 8.4.183 bumps Axelera Voyager SDK 1.8.0 → 1.9.0 and lets detection, instance/semantic segmentation, depth, pose, OBB, and classification all export, predict, and validate on Axelera Metis through yolo. NMS-free detection is the default. SDK 1.8 exports stay compatible; setup needs metis-dkms 1.7.0+ and PyTorch <2.15. Documented COCO128 throughput: 109 images/s at batch 32 vs 40 on SDK 1.8. Same release adds Intel CPU and NPU inference for LiteRT exports across those YOLO26 tasks.
Why it matters
One yolo export path now covers Metis plus Intel NPU for the whole YOLO26 task set — bid a Metis or Intel-NPU box, not a locked ALPR camera, when the city wants detect/segment/pose on the same silicon. Accelerated compute beats locked “AI cameras” (model swap vs 36–84 month refresh).
Source
GitHub v8.4.183 - 6
Key point
UNC Charlotte’s YUTELA (Yao, Ardabili, Kimiyaie, Tabkhi) is an edge–server–cloud AIoT stack for long-term, multi-site video anomaly detection. It replaces workstation nodes with NVIDIA Jetson Thor, batches cameras, uses shared memory and TensorRT, and keeps a separate anomaly model per camera retrained on a central GPU once enough recent pseudo-normal data lands. Eight-camera TensorRT lifts throughput 14.17 → 29.94 FPS; the Thor node beats the prior Ancilia workstation in 7 of 9 workload settings (up to 2.03× at extreme density). A 170-hour run on 31 existing CCTV cameras at four public sites produced 101 auto-triggered retrains using ~6.6% of one GPU; mean alert latency 6.8–15.4 s. Load tracked scene activity and input frame rate, not camera count.
Why it matters
This is a Thor + per-camera VAD pattern on cameras the city already owns — quote it when an MSP is asked to replace Flock with “analytics on the existing CCTV,” not another ALPR SaaS.
Source
arXiv:2610.26468 - 7
GeoForward: sparse LiDAR anchors feedforward 3D Gaussian splatting; lowest nuScenes cross-camera depth error
Key point
GeoForward (Shu et al.) conditions feedforward 3D Gaussian splatting on metric geometry instead of appearance-only matching. Sparse LiDAR points unproject to 3D, project across calibrated cameras as correspondence anchors, then a bounded visual search refines them; VGGT supplies global multi-view context. Separate heads predict metric depth and Gaussian attributes; a sparse 3D refinement keeps geometrically determined centers. Temporal reconstruction composes endpoint Gaussians with vehicle masks and motion priors. On nuScenes, GeoForward reports the lowest cross-camera depth-consistency error and Chamfer distance among compared methods while staying competitive on reconstruction quality.
Why it matters
LiDAR-anchored 3DGS is the occupancy / perimeter reconstruction path that does not wait on a city ALPR contract. Watch it for sites that already have a survey-grade or roadside LiDAR and want a camera-consistent 3D layer.
Source
arXiv:2610.23895 - 8
Key point
CGPLR (Iqbal, Ali, Tian, Zheng, Gou, Ou) is a confidence-guided pseudo-label refinement framework for unsupervised domain-adaptive person ReID. Sample reliability combines centroid compatibility, neighborhood consistency, separation from the strongest competing cluster, and temporal assignment stability. Cluster-adaptive thresholds then preserve, reassign, or defer labels; the loss is confidence-weighted for kept IDs, cautious for corrections, and label-free consistency for deferred samples. Results: 76.4% Rank-1 / 47.6% mAP on Market-1501 → MSMT17, and 95.1% Rank-1 / 87.7% mAP on MSMT17 → Market-1501. Code: github.com/Sundas058/CGPLR.
Why it matters
Cross-camera ReID still fails when a new site’s lighting and clothes shift the IDs. A preserve/reassign/defer gate is the control you want before an MSP turns on identity search across a new campus.
Source
arXiv:2610.25633 - 9
Key point
EMMA (Zhao, Kang, Li, Lin, Bai) is an event-driven multimodal memory system with multi-agent retrieval for long-video understanding that exceeds MLLM context windows. A dual-stream detector finds event boundaries from visual semantic drift and spoken transitions instead of fixed-length clips. Memory is a three-level hierarchy (raw video, clip-level, event-level). Retrieval decomposes the query, searches by tier, validates evidence, and answers. The authors report substantial gains on three long-video benchmarks vs clip-compression baselines. Code: github.com/BAI-LAB/EMMA.
Why it matters
Hours of perimeter video do not fit a VLM context window. Event-segmented memory is the retrieval pattern to test when a customer wants “ask the last shift” instead of another motion-search NVR.
Source
arXiv:2610.23892