Computer vision is the field of AI that extracts meaningful information — detection, tracking, counting and behaviour analysis — from camera images. It turns a security-camera stream into measurable events and metrics.
Work in this area
CX Teknoloji, founded by Abdulaziz Akyol, processes existing IP camera streams on GPU in real time: RTSP intake, hardware frame decoding, detection and tracking, a rule engine and event output. The articles share field experience from every link in that chain.
- 96.2% field accuracy in people counting, detection under 40 ms per frame
- 8–16 cameras per server; anonymous, privacy-by-design processing without face recognition
- Industrial scenarios such as forklift–pedestrian proximity, PPE detection and OEE
- Hands-on code guides with Python, YOLO, ONNX and TensorRT
Frequently asked questions (3)
Can video analytics run on existing security cameras?
Yes. ONVIF-compliant IP cameras that output H.264/H.265 are enough; the stream is pulled over RTSP and processed on a GPU server on site. No new cameras are needed — in most projects the real work is the network and camera placement.
Can video analytics comply with GDPR/KVKK?
Systems built without face recognition, without storing footage and producing only anonymous events and metrics greatly narrow the scope of personal-data processing. Notices, retention periods and access rights must still be part of the design.
Why does computer vision lose accuracy in the field?
Camera angle, night/IR images, where the counting line sits and tracking errors do not show up in the lab. Accuracy must be measured on site with real footage and the model calibrated with that data.
- Accelerating edge AI models: ONNX, TensorRT and INT8 quantizationArtificial intelligenceExport a PyTorch model to ONNX and speed it up with ONNX Runtime and TensorRT: FP16, INT8 calibration, accuracy loss and p50/p95 latency, with working code.
- Real-time object detection with Python and YOLO: RTSP camera to eventComputer visionReal-time detection from an RTSP camera with Python, OpenCV and YOLO: ByteTrack tracking, line counts, zone alerts and JSON events to MQTT, full script.
- Scaling GPU video analytics services with Docker and KubernetesSoftware developmentRunning a per-camera GPU analytics service on Docker and Kubernetes: NVIDIA Container Toolkit, GPU Operator, time-slicing/MIG, k3s, monitoring and updates.
- Video analytics on existing cameras: the chain from RTSP stream to eventComputer visionReal-time analytics on the IP cameras you already have: RTSP intake, GPU decoding, detection and tracking, a rule engine and event output, in five links.
- Edge or cloud? Choosing where to process in computer-vision projectsComputer visionProcess video on site or in the cloud? The decision we reached from bandwidth arithmetic, latency, data protection and three-year cost, and its exceptions.
- Getting people counting to 96%: field calibration notesComputer visionA model that scores 99% in the lab drops to 85% in the field. Camera angle, counting-line position, night data and measurement method close the gap to 96.2%.
- Data protection and video analytics: building analytics without face recognitionComputer visionFive questions legal teams ask about video analytics, answered by engineering: where video goes, what is stored, who has access, face recognition, retention.
- Forklift–pedestrian proximity detection: measurable safety rules in the warehouseComputer visionMeasuring forklift–pedestrian distance by camera: metres not pixels, a velocity vector not one frame, graded warnings not one alarm. The rule set behind it.
- Measuring OEE on the production line with camerasComputer visionTwo of OEE's three factors come from the PLC; the third and the reason for each stop hide in the camera. Extracting downtime and cycle time from video.
- Keeping detection models alive in the field: drift, retraining and dataset disciplineArtificial intelligenceTraining is the short part. The hard part is keeping a model accurate in 40 locations as season, light and layout change: drift, dataset and versioning.
Older technical notes (HTML5, PowerShell, Windows administration, Active Directory, HP/Aruba switches) are available in Turkish only. Türkçe →