Loading…
Loading…
22 production-ready AI analytics. Per-camera scene baselines for motion, audio, crowd, and tamper adapt to your site, and anomaly detection trains a baseline per camera, all on-premise, bounded, and operator-supervised. Cross-camera person journeys, ANPR, intrusion detection, directional counting, and more run on any stream with per-stream configuration and hardware-adaptive inference. Integrated with Radha AI Copilot for natural language control.
Each model is optimized for video analytics use cases with configurable parameters, threshold tuning, and per-stream activation.
Face detection, embedding extraction, and 1:N matching against enrolled galleries. Supports liveness detection and multi-angle recognition. Match quality depends on camera angle, lighting, and scene, so thresholds are calibrated per camera on your own footage.
Track individuals within each camera view using appearance-based feature vectors, then follow a person across the site. Version 1 cross-camera appearance search assembles a chronological journey with travel times, returning operator-reviewed candidates. A purpose-trained re-identification model for higher precision is in development.
Automatic Number Plate Recognition using a single global plate model rather than per-country templates, so it generalises across international plate formats. Infrared and visible-light camera support.
General-purpose object detection identifying 80+ object classes including vehicles, animals, luggage, and personal items. Real-time bounding box output with confidence scoring.
Crowd counting and occupancy estimation built on detection and tracking, rather than density-map regression. Configurable thresholds trigger alerts when defined capacity limits are exceeded.
Define restricted zones on any camera view. Alerts trigger when persons or objects enter, exit, or dwell within configured polygonal boundaries.
Directional line crossing detection for counting, flow analysis, and perimeter monitoring. Supports multiple lines per camera with per-direction triggering rules.
Detects presence or absence of personal protective equipment including hard hats, safety vests, gloves, and goggles. Designed for industrial and construction site compliance monitoring.
Pixel and vector based motion detection with configurable zones and sensitivity. Filters environmental noise like foliage and lighting changes to trigger only on meaningful movement.
Human body keypoint estimation for posture and activity analysis. Detects falls, raised hands, and other body configurations for safety and behavior monitoring.
Movement and dwell heat maps that visualize how people move through a space over time. Reveals high-traffic zones, bottlenecks, and dwell hotspots for operational planning.
Counts distinct individuals over a period using appearance features to avoid double-counting repeat visitors. Ideal for footfall analytics and occupancy trends.
Detects defocus, blackout, blinding, and obstruction so a sabotaged camera raises an alert instead of silently going dark. A per-camera scene baseline learns the day-to-night light cycle and suppresses false alarms.
Detects bags, packages, and objects left unattended beyond a configurable time, using owner-proximity and dwell analysis to cut false alarms in busy public spaces.
Detects two or more people passing through a secured door on a single credential (piggybacking), one of the most common access-control breaches, and alerts security in real time.
Measures queue length and average wait time per zone in real time, alerting when lines exceed your thresholds so staff can open a new counter before customers walk out.
Monitors each parking slot for a vehicle and reports overall lot occupancy in real time, alerting when the lot reaches capacity, using existing cameras with no per-space sensors.
Reports real vehicle speed from your cameras using a one-time per-camera calibration and alerts when a vehicle exceeds the limit you set, pairing with license-plate recognition for enforcement.
Learns what normal looks like for each individual camera and alerts on activity that deviates from it, with no rules, zones, or labels to configure, catching events your other analytics were never told to look for.
Spots visible flame and smoke on any camera and raises an alert in real time, so a fire is seen the moment it starts rather than when a ceiling sensor finally reacts. Runs on existing cameras with no dedicated fire or smoke sensors.
Detects visible firearms and knives on any camera and raises an immediate critical alert for operator review, confirming a weapon across several frames before alerting to suppress false alarms.
Listens to camera audio for gunshots, breaking glass, and screaming and raises an immediate critical alert for operator review, confirming a sound across several seconds before alerting to suppress false alarms.
Visylix automatically selects the optimal inference runtime based on available hardware. Deploy with or without GPUs.
Optimized inference on GPUs with FP16 and INT8 quantization. Batch processing across multiple streams maximizes GPU utilization and throughput.
Efficient CPU inference using hardware-accelerated optimization. Enables AI analytics on servers without dedicated GPU hardware, reducing deployment cost for smaller installations.
Enable, disable, and tune individual AI models on each video stream independently.
Enable face recognition on entrance cameras, ANPR on parking streams, and crowd detection in lobbies. Each stream runs only the models it needs.
Adjust confidence thresholds, minimum object sizes, and detection intervals per model per stream. Fine-tune sensitivity for each camera's field of view.
Define polygonal regions of interest within each camera view. AI models only process objects within configured zones, reducing false positives and compute load.
Configure where AI events are sent: WebSocket subscribers, webhook endpoints, recording triggers, or third-party integrations. Each model's events route independently.
Set per-stream FPS targets for AI processing. The engine balances inference load across available hardware to meet quality-of-service requirements.
New AI models under active development to expand Visylix's analytical capabilities.
Cross-camera appearance search and person journeys ship today as version 1, returning operator-reviewed candidates. A purpose-trained re-identification model for higher-precision cross-camera matching is in our training pipeline.
Adaptation in Visylix is specific and bounded, not a blanket claim. Motion, audio, crowd, and camera-tampering each maintain a per-camera scene baseline: ambient movement, noise ceiling, density, and the day-to-night light cycle, so foliage, traffic, and IR switches stop reading as events. Anomaly detection trains a baseline per camera at runtime, so it learns what normal looks like for that specific view. Where a source is noisy, the alarm-quality board proposes a threshold retune that an operator reviews and approves before anything changes. The remaining analytics are conventional trained models and do not self-improve in place. Nothing retrains itself from operator feedback, and nothing leaves the premises.
Motion learns each camera's ambient movement so foliage and traffic stop triggering it, audio learns its ambient noise ceiling, crowd density thresholds auto-calibrate, and camera-tampering learns the day-to-night light cycle so dark night scenes and IR switches no longer look like a blackout.
Anomaly detection trains a baseline for each camera at runtime, with no rules, zones, or labels to configure. It learns what normal looks like for that specific view and flags activity that deviates from it.
The alarm-quality board surfaces noisy sources and proposes a bounded threshold change. An operator reviews the suggestion and applies it. Visylix does not silently retune a safety detector on its own.
Acknowledging an alert, or dismissing it with a one-tap reason (false alarm, fog or steam, known person, duplicate), builds the alarm-quality record that retune suggestions are drawn from. The feedback informs the suggestion; it does not retrain a model by itself.
An opt-in, human-approved review queue lets a site's own reviewed examples feed a future model update, with eval-before-swap so a new model only ships if it beats the old one. Nothing trains automatically.
All baseline calibration happens on your hardware and survives restarts. Video data never leaves your building, nothing is sent to any cloud, and it works in air-gapped environments.
Draw custom polygons and virtual lines on any camera view. Assign AI models per zone, filter events by region, and count directional crossings with speed estimation.
Hardware-adaptive inference automatically selects GPU or CPU based on available hardware. Scale from small test deployments to enterprise production environments.
Optimized NVIDIA CUDA inference for maximum throughput. Recommended for production deployments requiring real-time AI processing at scale.
Efficient CPU-only inference for testing environments and small deployments. No dedicated GPU hardware required, reducing deployment cost.
Automatic GPU detection and configuration. The engine automatically selects GPU or CPU inference based on available hardware.
See how Visylix AI integrates with the VMS Core Engine and multi-protocol streaming to deliver actionable intelligence.
Visylix includes 22 production-ready AI analytics: face recognition, per-camera person tracking, ANPR (license plate recognition), object detection, crowd detection, pose estimation, PPE detection, heat map generation, motion detection, unique person counting, intrusion detection, line crossing detection, camera tampering detection, abandoned object detection, tailgating detection, queue and wait-time analytics, parking occupancy, speed estimation, anomaly detection, fire and smoke detection, weapon detection, and audio event detection. Motion, audio, crowd and tamper analytics maintain per-camera scene baselines, and anomaly detection trains a baseline per camera at runtime; the rest are conventional trained models. Detection accuracy depends on camera angle, lighting and scene, so we validate each analytic on your own footage during the proof of concept.
Yes. Visylix AI Analytics features hardware-adaptive inference supporting both GPU and CPU platforms, with per-stream AI configuration for maximum flexibility.
Adaptation in Visylix is specific and bounded, not a blanket claim. Motion, audio, crowd and tamper analytics maintain per-camera scene baselines that adjust to gradual lighting, day and night cycles, and weather. Anomaly detection trains a baseline per camera at runtime, so it learns what normal looks like for that view. Where thresholds drift, operators receive retune suggestions they review and approve before anything changes. Nothing retrains itself from operator feedback without approval, and the remaining analytics are conventional trained models that do not self improve in place.
Yes. Visylix supports polygon-based detection zones with per-zone AI model assignment, named zones for clear alert identification, and per-zone statistics including detection counts and dwell times. You can also configure virtual tripwires with directional line crossing, object type filtering, per-line crossing statistics, and speed estimation.
Yes. In GPU mode with NVIDIA CUDA, Visylix supports GPU-accelerated real-time inference for object detection, recommended for production deployments. CPU-only mode runs at a reduced frame rate and stream count, ideal for testing and small deployments; we size it on your own hardware during the proof of concept. The engine automatically detects available hardware and selects the optimal inference mode.