Sanchara Traffic Systems builds two ITS products that run on the cameras agencies already operate, both currently in live R&D pilots. FreeACT detects freeway incidents in real time. InterACT turns intersection video into calibrated turning movements, queue, delay, and conflict analytics.
InterACT runs on recorded clips or live RTSP, then outputs engineering-grade metrics: turning counts, queue and delay, calibrated speed, and conflict analytics in analysis-ready CSVs.
Same engine, two specialized analytics layers, one for freeway incident response, one for intersection performance.
Real-time AI incident detection across the cameras agencies already operate. Single-batch GPU inference, ~28s detection latency, automated video evidence with 45s pre-incident buffer, full TMC lifecycle, self-learning false-positive suppression.
Engineering-grade intersection data from any traffic camera. Calibrated mph from a 4-point homography, ByteTrack with occlusion recovery, per-lane queue, delay, and headway, plus FHWA 32-point conflict-grid analysis.
DOTs operate hundreds of cameras and a handful of operators. The math does not work. Sanchara Traffic Systems lets the cameras you already operate do the watching, the counting, and the alerting.
Manual freeway-incident detection runs on the order of several minutes, with field studies reporting times in the 5–15 minute range[1]. Each additional minute of incident duration raises the likelihood of a secondary crash by roughly 2.8%[2]. ALDOT I-20/59 today operates 44 cameras into a small operator pool; published ergonomics research caps effective live monitoring at 9–16 simultaneous feeds[3].
DOTs already operate the camera infrastructure. No new sensors, no trenching, no hardware procurement. The system runs on existing RTSP feeds.
InterACT replaces manual turning movement counts with calibrated, per-second video-derived analytics. Same camera, same hour-of-day, repeatable across weeks.
The same engineering decisions show up in FreeACT and InterACT, because they're what make a single-GPU system handle dozens of live feeds in real-world R&D pilots.
All cameras, one batch, every cycle.
FreeACT and InterACT both run every connected RTSP camera through a single YOLO12 + TensorRT batch on one GPU. Sub-2s cycle times across 44 cameras on an RTX 3080 Ti, 4 simultaneous videos in InterACT batch mode, dynamic resize on OOM.
Share the reader, not the connection.
FreeACT writes pre-annotated JPEG frames to a tmpfs RAM disk and reuses that buffer for clip recording, no second RTSP connection per incident. The same reader serves the live MJPEG dashboard and the post-trigger MP4 encoder.
Alert fatigue decreases over time.
Per-camera evidence multipliers (+0.15/dismiss, decay 0.02/day) and permanent spatial suppression zones learn from operator feedback. Zones are gated on movement so vehicles that drove in then stopped are never blocked. 14-day auto-expiry.
Same software. Laptop to workstation.
Detects GPU, VRAM, CPU cores, NVENC, and NIC at startup. Picks batch size 10–102, resolution 480–800px, FP16 vs FP32, and decode FPS automatically. Falls back to CPU if no CUDA. Same binary runs on a 4 GB laptop and a 16 GB workstation.
The system is built, tested, and running today on I-20/59 and the University of Alabama campus. We can demo it on your cameras next.