Visylix delivers intelligent video analytics through 22 specialized AI models, from face recognition to abandoned object detection, plus Radha AI Copilot for natural language control. Model adaptation is per camera and bounded, on-premise, and operator-supervised. Every model runs on your own hardware, GPU-accelerated where an accelerator is available and CPU-only where it is not.
Each model is purpose built and optimized for its domain, delivering best in class accuracy and performance.
Enterprise-grade facial detection and identity verification with anti-spoofing, multi-face tracking, and fast database matching against large-scale identities.
Per-camera tracking that maintains consistent identities within each camera view, plus cross-camera appearance search (v1): pick a person and Visylix assembles their journey across your other cameras as operator-reviewed candidates.
Automatic number plate recognition (ANPR/ALPR) reading global plate formats with an India-optimized pipeline, night-vision IR capability, and moving-traffic capture.
Real-time body skeleton detection tracking 17 keypoints per person for safety monitoring, movement analysis, fall detection, and gesture-based interaction.
Enterprise-grade computer vision identifying and localizing 80+ object classes with custom class support via transfer learning and persistent tracking across frames.
AI-powered density estimation and people counting using regression-based density maps, accurate from sparse to ultra-dense environments where traditional detection fails.
Specialized PPE detection for industrial environments identifying hard hats, vests, gloves, glasses, and boots with real-time compliance scoring and violation alerts.
Transforms person-detection data into color-coded spatial overlays revealing movement patterns, dwell times, traffic flows, and engagement hotspots.
Intelligent motion analysis using adaptive background modeling and shadow suppression with configurable zones, tripwires, and directional rules for precise alerting.
De-duplicated visitor counting that combines person detection and tracking with in-memory de-duplication for accurate footfall analytics, not inflated headcounts.
Zone-based unauthorized entry detection using AI-powered virtual perimeters with instant alerts, eliminating the need for physical sensors or laser barriers.
Directional boundary monitoring that counts and alerts when people or vehicles cross defined virtual lines, with configurable direction rules and counting analytics.
Automatically detects camera sabotage such as defocus, blackout, blinding, and obstruction, so a blinded or covered camera raises an alert instead of silently going dark.
Detects unattended bags, packages, and objects left in a monitored area, using owner and dwell-time analysis to cut false alarms in busy environments.
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. Runs on any camera covering a door, turnstile, or mantrap with no extra hardware.
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. Built for retail, banking, and transport hubs.
Monitors each parking slot for a vehicle and reports overall lot occupancy in real time, alerting when the lot reaches capacity. Uses existing overhead or wide-area cameras with no per-space sensors, and ignores cars merely driving past.
Reports real vehicle speed from your cameras using a one-time per-camera calibration and alerts when a vehicle exceeds the limit you set. Works alongside license-plate recognition so an over-speed event can carry the vehicle plate, with no radar or in-road sensors.
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. A semantic per-camera memory built from the scene's own visual embeddings catches unusual activity and out-of-place objects, not just raw pixel change.
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 the cameras you already have, with no dedicated fire or smoke sensors, and can be limited to the areas that matter.
Detects visible firearms and knives on any camera and raises an immediate critical alert for operator review. It confirms a weapon across several frames before alerting to suppress false alarms, and runs on the cameras you already have with no special hardware.
Listens to camera audio for gunshots, breaking glass, and screaming and raises an immediate critical alert for operator review. It confirms a sound across several consecutive seconds before alerting to suppress false alarms, and works on any camera that has a microphone.
Our models are trained, tested, and deployed for production grade video intelligence.
Battle tested models deployed across enterprise environments with proven reliability.
GPU-accelerated real-time inference for object detection. CPU-only mode is available for testing and small deployments.
Optimized model variants for edge compute devices with hardware-accelerated inference.
Per camera model assignment with adjustable sensitivity, zones, and scheduling.
Adaptation is per camera, on-premise, bounded, and always operator-supervised. Motion, audio, crowd, and camera-tampering detection build per-camera scene baselines, and anomaly detection trains a per-camera model. For the threshold-based detectors, Visylix surfaces retune suggestions an operator approves. The remaining models are conventional trained models that do not adapt in place. Nothing retrains itself from operator feedback, nothing leaves your building, and every adaptation is reversible.
Motion learns each camera's ambient movement so foliage and traffic stop triggering it, audio learns the ambient noise ceiling, crowd thresholds auto-calibrate to each scene, and camera-tampering learns the day-to-night light cycle so dark nights and IR switches stop firing false blackout alarms.
Anomaly detection trains a model on footage from that camera itself rather than shipping a fixed rule, so what counts as normal is learned from that scene. Training runs on-premise and the learned baseline is reversible.
When operators dismiss alerts from a noisy source, the alarm-quality board surfaces a retune suggestion. An operator approves it before anything changes, the step is bounded, and for safety detectors (fire, weapon, PPE) an approved retune can only make them more conservative, never less.
Anomaly detection builds a per-camera memory from the scene's own visual embeddings, catching unusual activity and out-of-place objects rather than raw pixel change. It keeps separate day and night memories, so nightfall is not treated as an anomaly.
Acknowledging an alert, or dismissing it with a one-tap reason (false alarm, fog or steam, known person, duplicate), scores the source on the alarm-quality board. A one-click apply-retune is then offered to the operator, who decides whether to raise that source threshold by a bounded step.
An opt-in, human-approved fine-tune queue lets your own reviewed examples improve the detector models, with eval-before-swap so a new model only ships if it beats the old one. Nothing trains automatically, nothing leaves the premises, and all state survives restarts.
Follow a person or a vehicle across your site, not just within one camera view. Available in plain language through Radha (“where did this person go?”, “track that vehicle across the site”).
Pick a person from any detection and Visylix finds appearance matches on your other cameras within a time window, assembling a chronological path with travel times between cameras. It is appearance-based and returns candidates for operator review, not automatic identity assertions. A purpose-trained re-identification model for higher precision is in our training pipeline.
Because a plate is an exact identifier, vehicle journeys are precise: the same plate seen at camera A then camera B is stitched into one cross-camera path with the travel time between them, for enforcement, investigation, and site-flow analysis.
Schedule a live demo to see how Visylix AI models perform on your video feeds.
In a bounded, on-premise, operator-supervised way. Motion, audio, crowd, and camera-tampering detection build per-camera scene baselines, and anomaly detection trains a per-camera model from the visual embeddings of that scene. For the threshold-based detectors, Visylix surfaces retune suggestions on the alarm-quality board that an operator reviews and approves; the system does not retrain itself from operator feedback, and for safety detectors an approved retune can only make them more conservative, never less. The other models are conventional trained models that do not adapt in place. Nothing leaves the premises, and every adaptation is reversible and survives restarts. We do not claim a false-positive reduction figure.
Yes, as a v1 cross-camera appearance search. Pick a person from any detection and Visylix finds appearance matches on your other cameras within a time window and assembles a chronological journey with travel times between cameras. It is appearance-based (visual similarity) and returns candidates for operator review, not automatic identity assertions. A purpose-trained re-identification model for higher precision is in our training pipeline. Vehicle plate journeys are exact, since a plate is a precise identifier.
No. Visylix AI is fully on-premise. All inference, learning, scene baselines, face templates, and cross-camera search run on your own hardware and work in air-gapped deployments. Your video and model adaptation never leave your building.
By default Visylix Tailgating Detection counts how many people pass through a secured door during an access event, from video alone, and flags group crossings. Opt in to badge-to-body correlation with a supported access-control system (HID, Lenel S2, AMAG, Brivo, Genea, Bosch) and Visylix compares how many people crossed the door with how many badges were presented, alerting only when more people entered than credentials, which is the true tailgating signal.
Yes. Queue and Wait-Time Analytics reports the current line length, the live service rate (how fast the queue is actually clearing), and a predicted wait time for someone joining the queue now, computed from that clearing rate, alongside average and maximum wait. It alerts when a line grows past your threshold so staff can open another counter before customers leave.
Yes. Line Crossing Detection produces in, out, and net people counts per line with hourly breakdowns on a dashboard, not just individual crossing events, the directional counting product that retail and facilities teams expect.