Smart cities and traffic
Helmet and traffic violation detection on highways and junctions, and crowd monitoring in public areas.
Deep-learning video analytics
AI Video Analytics
Purpose-trained models watch live and recorded camera video for security, safety, traffic and operational events, and raise alerts in real time.
What AI Video Analytics does
It analyses camera video, turns what it finds into structured events and alerts the people who need to act.
A control room usually has more cameras than its operators can watch at once. Events can be missed live and found only later, in playback.
AI Video Analytics runs purpose-trained deep-learning models on the camera streams. The models detect people, objects, vehicles and behaviours in real time. Each detection is checked against the rules set for that camera, and what qualifies is raised as an alert.
Detections are also kept as metadata, so recorded video can be searched by what is in it and not only by time.
- Video
- Live and recorded, from existing IP cameras
- Alerts go to
- Video management software (VMS), SMS, email or webhook
- Runs on
- Edge devices, GPU servers or cloud instances
Analytics modules by purpose
Each module is a deep-learning model trained for one purpose, with its own rules engine. Modules are chosen per camera, so one site can run several.
- Security
-
- Intrusion detection. A tripwire crossed or a restricted zone entered.
- Loitering detection. A person who stays in a sensitive zone beyond the set time.
- Object detection. Abandoned objects, missing objects and unauthorised removal.
- Crowd monitoring. Crowd density, flow and suspicious behaviour.
- Face recognition. Faces matched in real time against staff, VIP or watchlist databases. See the Face Recognition page.
- Safety
-
- Fire and smoke detection. Flames and smoke that are visible on camera.
- PPE compliance. Missing personal protective equipment (PPE): hard hat, vest, gloves or mask.
- Traffic
-
- Helmet and traffic violations. Riders without a helmet, triple riding and wrong-way driving.
- Vehicle analytics. Vehicle class, direction, dwell and parking, and under-vehicle inspection.
- Operations
-
- People counting. People moving in and out at gates, docks and entrances, counted in both directions.
- Heatmap and dwell time. Where people move in a space and how long they stay.
- Video synopsis. Hours of footage condensed for review, with search by class, colour, direction or time.
What a detection looks like
The analytics draw what they find on the video frame: a box round the detection and a text label. These three frames are analytics output from real footage.
Missing gloves on an assembly line
This overhead frame shows the gloves check of the PPE compliance module. Orange boxes mark the bare hands of two workers. A marked frame shows what was flagged and where in the scene.
In and out counting at a loading dock
A counting line is set across the camera view. Each crossing adds to an in counter or an out counter, shown here on a night view of a loading dock.
A helmet violation on a city highway
A rider without a helmet is marked and labelled on a highway camera. For traffic enforcement the module can be paired with automatic number-plate recognition (ANPR), so that a violation becomes a plate-linked evidence record.
Tuning each module to its camera
Zones, schedules and confidence thresholds are set for each camera. With scene calibration, they are the controls used to keep false alarms down.
- Zones and lines
- A restricted zone, a tripwire or a counting line limits a rule to the part of the scene where it applies.
- Schedules
- A schedule limits a rule to the hours when it applies. For example, an entrance may be counted during opening hours and watched for intrusion after closing.
- Confidence thresholds
- The threshold sets how confident the model must be before an alert is raised. A higher threshold gives fewer false alarms and may miss more events. A lower one does the opposite.
- Scene calibration
- Scene calibration adjusts a module to the view of one camera, to suppress false alarms. It is part of scale-out on site, and optimisation continues after handover.
- Custom model training
- Where a site needs to detect an object that the standard modules do not cover, a model can be trained for it.
Field conditions
Models are trained and hardened for glare, dust, monsoon rain, low light, dense crowds and mixed camera quality. Results still depend on camera position and lighting, which is why a rollout starts with a site survey and camera audit.
Edge, on-premise or cloud
The same modules run in three places, and one estate can mix them. Network and GPU sizing follows the site survey.
| Option | Suits | Where the models run |
|---|---|---|
| Edge | Sites with poor or no connectivity. Detection keeps working when the link is down. | On the camera or on an edge box |
| On-premise | Sites where video must stay on the network, including air-gapped sites. | On GPU servers on your own network |
| Cloud | Multi-site estates that need central dashboards, updates and storage. | On cloud instances |
Video in, alerts and data out
The analytics take video from the cameras a site already has and send events to the systems its teams already use.
- Cameras and recorders
- Video comes from existing IP cameras and network video recorders (NVRs) over ONVIF and RTSP.
- VMS Enterprise
- Alerts appear in the VMS Enterprise console, beside live view, recording and playback.
- An existing VMS
- The analytics can be bridged into a VMS or command-centre console that the site already runs.
- Command and Control
- In Command and Control, alerts are shown on the map and handled through escalation workflows.
- SMS and email
- Alerts reach phones and inboxes through SMS and email gateways, with escalation rules and schedules.
- REST API and webhooks
- The REST API and webhooks push events, metadata and alerts to dashboards, and to security information and event management (SIEM) and physical security information management (PSIM) systems.
Where it is used
Each sector uses a different mix of modules. These pages describe the typical deployments.
Defence and paramilitary
Intrusion detection with tripwires and restricted zones on site perimeters, running on air-gapped networks.
Airports and transport
Unattended-baggage alerts in terminals, and crowd monitoring on platforms and concourses.
Retail and malls
Heatmaps, dwell time and people counting in stores, and queue-length alerts at billing counters.
Industrial and warehousing
PPE compliance on production lines, danger-zone alerts on shop floors, and in and out counting at docks.
Case study: airport terminal operations
Unattended-baggage and crowd analytics on server GPUs, with alerts sent to the terminal control room.
Datasheet
AI Video Analytics
Deep-learning video analytics software. Purpose-trained models analyse live and recorded camera video, apply per-camera rules and send real-time alerts.
| Product | AI Video Analytics, deep-learning video analytics software |
|---|---|
| Models | Purpose-trained deep-learning models, each with a rules engine |
| Video input | Live and recorded video from existing IP cameras and network video recorders, over ONVIF and RTSP |
| Security modules | Intrusion detection (tripwire, restricted zones); loitering detection; object detection (abandoned objects, missing objects, unauthorised removal); crowd monitoring (density, flow, suspicious behaviour); face recognition |
| Safety modules | Fire and smoke detection; personal protective equipment (PPE) compliance (hard hat, vest, gloves, mask) |
| Traffic modules | Helmet and traffic violations (no helmet, triple riding, wrong way); vehicle analytics (classification, direction, dwell and parking, under-vehicle inspection) |
| Operations modules | People counting (in and out, both directions); heatmap and dwell time; video synopsis (search by class, colour, direction or time) |
| Per-camera settings | Zones, schedules and confidence thresholds |
| False-alarm control | Scene calibration |
| Custom models | Model training for site-specific objects |
| Alert routing | Video management software (VMS), SMS, email or webhook |
| API | REST API and webhooks for events, metadata and alerts |
| Deployment | Edge devices, on-premise GPU servers or cloud instances |
| Sizing | Determined during the site survey |
Videolytical Systems Pvt. Ltd., 277, C Block, Sector 63, Noida, Uttar Pradesh, India. Sales and demos: sales@videolytical.com, +91 97160 05441.
Questions buyers ask
Do we need new cameras?
Usually not. The analytics are added through edge boxes or GPU servers that take the streams of your existing IP cameras and recorders over ONVIF and RTSP. Existing cameras are reviewed in the site survey and camera audit.
Does it work without an internet connection?
Yes. The models run offline on edge devices or on-premise GPU servers. Cloud is an option for multi-site estates. Detection and alerting do not depend on it.
Can it detect something specific to our site?
Yes. Custom models can be trained for site-specific objects. Which models are selected and which are trained is decided in the AI model design step of how we deliver.
How are false alarms kept down?
Rules are limited by zone, schedule and confidence threshold, and scene calibration is used to suppress false alarms. A pilot on a small set of cameras shows the results on your own scenes before scale-out. Our guide to reducing false alarms explains the controls.
Who installs and supports the system?
Videolytical delivers with OEMs (camera and hardware makers), security solution providers and system integrators. Delivery includes testing and calibration, and optimisation continues after handover.
See the analytics on your own footage
Arrange a live demo, or send a sample clip and we will run the modules you choose on the scenes you care about.