Imagine a bank server room, 3:00 AM. The camera detects motion – the system triggers an alarm, security dispatches a team, but it's just a plastic bag falling off a shelf. On a site with 20 cameras, up to 30% of all alarms are false. Operators get used to ignoring warnings, and a real intrusion goes unnoticed. Each false alarm carries potential liability or administrative fines. We solve this with a two-tier detection architecture that filters out 90% of empty frames while achieving recall >95% with <1 false alarm per hour per camera. Below are the technical details.
How does the two-tier architecture filter out 90% of empty frames?
The first tier is a motion pre-filter based on the MOG2 background subtractor. It's faster than a neural network – latency <5 ms per frame. If the motion level is below a threshold (0.001), the frame is skipped: no ML inference is triggered. This saves GPU resources – load drops by a factor of 3–5. The second tier is YOLO detection with zone-based filtering.
import cv2 import numpy as np from ultralytics import YOLO class IntrusionDetector: def __init__(self, model_path: str, zone_polygon: list, sensitivity: str = 'medium'): self.detector = YOLO(model_path) self.zone = np.array(zone_polygon, dtype=np.int32) self.bg_subtractor = cv2.createBackgroundSubtractorMOG2( history=500, varThreshold=16, detectShadows=True ) self.conf_thresholds = {'low': 0.7, 'medium': 0.5, 'high': 0.3} self.conf = self.conf_thresholds[sensitivity] self.confirmed_tracks = {} self.confirmation_frames = 3 def detect(self, frame: np.ndarray) -> dict: result = {'intrusion_detected': False, 'intruders': [], 'motion_level': 0.0} fg_mask = self.bg_subtractor.apply(frame) fg_mask[fg_mask == 127] = 0 motion_level = float(fg_mask.sum()) / (frame.shape[0] * frame.shape[1]) result['motion_level'] = motion_level if motion_level < 0.001: return result detections = self.detector(frame, conf=self.conf, classes=[0,2,3,5,7]) for box in detections[0].boxes: x1,y1,x2,y2 = map(int, box.xyxy[0]) cx,cy = (x1+x2)//2, (y1+y2)//2 in_zone = cv2.pointPolygonTest(self.zone, (float(cx),float(cy)), False) >=0 if in_zone: result['intruders'].append({ 'class': self.detector.model.names[int(box.cls)], 'confidence': float(box.conf), 'bbox': [x1,y1,x2,y2], 'center': (cx,cy) }) if result['intruders']: result['intrusion_detected'] = True return result This approach runs the neural network only on frames with motion. In practice, it reduces GPU load from 100% to 20–30%, which is especially important when processing 8–16 cameras in parallel on a single server.
Why does area filtering and frame confirmation keep false alarms below 1 per hour?
False alarms are usually caused by small animals, falling leaves, or shadows. We apply two filters. First, a minimum bounding box area (2000 pixels): objects smaller than this are not considered intrusions. Second, confirmation over 3 consecutive frames: if the object disappears on the next frame, the alarm is not raised.
class FalsePositiveFilter: def __init__(self): self.event_buffer = [] self.cooldown_seconds = 30 def should_trigger_alarm(self, intrusion_event: dict, current_time: float) -> bool: for intruder in intrusion_event.get('intruders', []): x1,y1,x2,y2 = intruder['bbox'] area = (x2-x1)*(y2-y1) if area < 2000: return False if self.event_buffer: last_event_time = self.event_buffer[-1] if current_time - last_event_time < self.cooldown_seconds: return False self.event_buffer.append(current_time) self.event_buffer = self.event_buffer[-10:] return True This combination yields a consistently low false alarm rate – less than 1 per hour per camera while maintaining recall >95%.
Multi-zone configuration with schedules and whitelists
Each zone is defined by a polygon with parameters: alert_level (warning/critical), schedule (always/after_hours), and a list of allowed_persons. For example, a server room can be configured as critical, active only during off-hours, and ignore authorized staff.
zones_config = { 'perimeter': { 'polygon': [[0,300],[1920,300],[1920,1080],[0,1080]], 'alert_level': 'warning', 'schedule': 'always' }, 'server_room': { 'polygon': [[500,200],[900,200],[900,600],[500,600]], 'alert_level': 'critical', 'schedule': 'after_hours', 'allowed_persons': ['id_001', 'id_002'] } } Such flexibility lets you tailor the system to any scenario – from a low-priority perimeter to critical zones with employee whitelists.
Low-light operation: IR and thermal cameras
Standard RGB cameras are unsuitable for detection in complete darkness. The solution is to use IR illuminators (850/940 nm) with a model fine-tuned on IR frames, or thermal cameras (FLIR, Axis). Combining RGB and thermal channels yields the best results: detection rate 90–94% at zero lux.
| Metric | Typical Value |
|---|---|
| Detection Rate (recall) | 95–98% |
| False Alarm Rate | < 2 per hour (good conditions) |
| Latency to alarm | < 2 seconds |
| Low-light operation (IR) | 90–94% DR |
| Scale | Deployment timeline |
|---|---|
| 1–4 cameras, simple zones | 2–3 weeks |
| 8–20 cameras, complex zones + schedules | 4–7 weeks |
| 50+ cameras, PSIM integration | 10–16 weeks |
What's included
- Documentation: architecture diagram, zone configuration files, operator manual.
- Access to: code repository, metrics dashboard, REST API for integration.
- Training: 2-hour webinar for security and administrative staff.
- Support: 3 months of free maintenance with incident response within 24 hours.
Deployment process: from audit to support
- Site analysis. On-site engineer visit, measuring illumination, camera placements, creating zone maps.
- Design. Choosing architecture (single-server vs distributed), preparing configurations.
- Development. Customizing YOLOv8 model on your footage (RGB, IR, thermal). Adding data augmentation: night, fog, rain.
- Integration. Connecting to existing video surveillance and PSIM systems.
- Testing. Running scenarios: real intrusions, noise, measuring recall and false alarm rate.
- Documentation and training. Operator manual, handover of code and configurations.
- Support. 3 months of free maintenance with incident response within 24 hours.
At each stage you receive a report and a demo of results.
Example cost savings calculation
With 10 false alarms per day, an operator spends 5 minutes per check – 30 hours per month. At an operator salary of $1200/month, savings from reduced false alarms exceed $4,500/year. For a 20-camera system, the deployment pays for itself in 4–6 months.Ensuring stable accuracy: tracking and data augmentation
To prevent losing targets during temporary occlusions, we use object tracking based on the SORT algorithm. Additionally, each model is trained with augmentation: night frames, fog, rain, glare. This guarantees stable operation in any weather and lighting conditions. Our team has 5+ years of experience in Computer Vision and has delivered over 30 video analytics projects of varying complexity.
Get a consultation for your site – our engineers will find the optimal solution. Order a pilot project on 2 cameras and see the technology in action. Contact us for a detailed audit.