Use case
Intrusion / Trespassing
Virtual perimeter zones trigger instant on-device alerts when a person or vehicle crosses a restricted boundary, day or night. Inference results are pushed to the cloud within seconds of detection, enabling immediate dispatch without waiting on a human reviewing a feed.
Where This Can Be Deployed
The types of locations and operational environments where this solution could be installed.
Industrial compounds
Industrial parks
Provides round-the-clock perimeter awareness across large industrial compounds.
Construction sites
Water-treatment facilities
Flags unauthorised access to active or high-value construction sites.
Utility infrastructure
Telecommunications sites
Helps protect unmanned utility infrastructure from tampering or theft.
Warehouses
Construction zones
Supports after-hours monitoring of warehouse perimeters without on-site staff.
Restricted government facilities
Storage yards
Adds an automated layer of perimeter awareness to restricted facilities.
Solar farms and remote assets
Restricted perimeter areas
Extends monitoring to remote assets that are impractical to patrol regularly.
From a General Model to Site-Specific Intelligence
The initial model provides general object and event detection capabilities. Its performance improves as it is trained and validated using footage from the actual deployment environment, including local lighting, weather, camera angles, traffic patterns and real operational events.
Stage 1 — Baseline Model
General detection model
- Uses a pre-trained general model
- Detects common objects and activities
- Provides an initial baseline
- Has not yet learned the site's unique conditions
- Performance may be affected by camera angle, rain, glare, shadows, occlusion and local behaviour
Stage 2 — Site Learning
Continuous validation and tuning
- Collects representative footage from the deployment environment
- Reviews correct detections, missed events and false detections
- Adds verified event examples to the training dataset
- Tunes detection thresholds and event rules
- Tests performance across day, night and adverse-weather conditions
Stage 3 — Three-Month Site-Adapted Model
Site-specific intelligence
- Better adapted to the actual camera perspective
- Improved recognition of locally relevant events
- Reduced avoidable false alerts
- Better performance under recurring site conditions
- More useful alerts for the operational team
After approximately three months of representative footage, event validation and model tuning, the system can become significantly better adapted to the deployment environment. Actual improvement depends on footage quality, event frequency, annotation quality, camera positioning and operating conditions.
How Site Training Improves the Model
- Malaysian weather, including heavy rain and changing sunlight
- Night-time lighting and low-visibility conditions
- Local vehicle types and road behaviour
- Camera-specific viewing angles
- Site-specific restricted zones
- Typical pedestrian and crowd movement
- Common false-alert sources
- Rare but operationally important events
Verified footage and labelled events are reviewed and incorporated into controlled model-training and validation cycles.
Expected Model Improvement
A transparent comparison between a general model and one adapted to this deployment — with no numbers claimed until they are backed by real evaluation data.
General Baseline Model
- General-purpose detection
- Limited understanding of the site
- Initial benchmark required
- More sensitive to unfamiliar conditions
Site-Adapted Model
- Trained using verified local footage and events
- Tuned for the installed camera perspective
- Better adapted to local operational conditions
- Designed to reduce unnecessary alerts
Detection performance must be measured using a validated test set from the intended deployment environment. Results vary according to use case, camera position, footage quality, event frequency and environmental conditions.
Use-Case-Specific Detection Metrics
The metrics that matter most for this use case. Only metrics backed by real test data will show a measured value — everything else is shown as a neutral status until then.
Precision
Recall
Missed-event rate
Day-versus-night performance