CCTV & AI Video Analytics Glossary: Key Terms ExplainedLearn 24 essential CCTV and AI video analytics terms with practical notes, including RTSP, ONVIF, Hype-Model, on-premise setups, and more.

RTSP is a protocol used to stream video from IP cameras over a network in real time. It supports main and substreams in compression formats like H.264 or H.265 to optimize bandwidth.

In practice, camera feeds can connect directly to analytics devices or VMS via an RTSP URL. Substreams are often used for applications requiring lower bandwidth or reduced latency.

ONVIF is a standardized protocol ensuring interoperability between IP cameras and recorders from different brands. Profile S handles basic video streaming, Profile T supports advanced video and analytics, and Profile M covers device management.

Unlike RTSP, ONVIF also includes device discovery, control, and configuration management, simplifying system integration and scalability.

RTSP is solely a video streaming protocol, whereas ONVIF provides a comprehensive framework for both streaming and device control/configuration. ONVIF can use RTSP streams but adds camera management capabilities.

Therefore, ONVIF compatibility facilitates broader ease of integration, device setup, and monitoring.

An NVR stores and plays video streams from IP cameras via RTSP or ONVIF. Recordings are maintained on disk and accessed through VMS software.

Some NVRs include built-in analytics modules, but advanced AI processing typically requires centralized GPU servers or edge AI solutions.

VMS is software that centrally manages multiple camera feeds, offering recording, monitoring, and alarm management. It supports RTSP and ONVIF connections for flexible integration.

VMS provides multifunctional interfaces like alarm pop-ups, event marking in recordings, and data export to ERP/MES systems.

H.264 and H.265 are industry-standard codecs reducing bandwidth and storage needs by compressing video data. H.265 is about 50% more efficient than H.264, especially for high-resolution streams.

Both codecs are commonly used for high-quality, low-latency streams in AI analytics.

The main stream transmits high-resolution, high-frame-rate video, while the sub stream uses lower resolution and frame rates to save bandwidth.

AI analytics typically prefers the main stream for detail and accuracy, whereas sub streams are suitable for multi-camera monitoring in VMS.

FPS (Frames Per Second) indicates how many images are processed each second. Stable and sufficient FPS is critical for AI object detection and behavior analysis.

Typically, streams over 10-15 FPS suffice for AI, but faster-moving objects may require 20 FPS or higher.

Resolution defines the pixel count in an image, while pixel density affects clear recognition of objects without confusion. Higher resolution and pixel density improve detection of small objects.

Cameras of 2 MP and above are preferred for quality AI analysis, though 4K or higher may be used depending on the application.

ROI refers to prioritized or critical areas within an image for processing. AI algorithms can extract more details or scan these areas more frequently.

Using ROI reduces computational load and minimizes latency caused by processing irrelevant data.

Object detection identifies targets such as people, vehicles, or packages in video streams. This is a core function of AI models.

Model accuracy and real-time performance depend on factors like lighting, resolution, and FPS.

Classification determines the category of detected objects. For example, vehicles may be classified as cars, trucks, or motorcycles.

This enhances AI models by providing more detailed and actionable results.

Pose estimation analyzes human figure movements and postures, while hand tracking monitors hand positions and motions.

These analyses are important for industrial applications in safety and human-robot interaction.

OCR technology converts text and numbers in images into readable data. It is used for license plate, barcode, or label reading from camera footage.

AI-enhanced OCR uses advanced algorithms robust to font variations and lighting conditions.

VLM combines visual data with natural language processing capabilities. It analyzes and interprets images, then interacts through text or commands.

It can be used in industrial systems for data interpretation and reporting.

"Hype-Model" is Hype Vision’s AI model family, including modules for object detection, classification, pose/hand tracking, and OCR.

These models are continuously updated, optimized for high accuracy and low latency, and compatible with cameras independent of brand.

Edge AI performs AI processing on or near the device (camera) where data is generated. This allows real-time processing and rapid response without sending data to central servers.

It is preferred in low-latency-critical applications, reducing network bandwidth and enhancing data security.

Centralized servers equipped with powerful GPUs accelerate AI analysis for many camera streams. This architecture provides high processing power but requires consideration of bandwidth and latency.

The number of GPUs needed depends on camera count, resolution, FPS, and model complexity.

The required GPU count for processing 100 cameras varies based on resolution, FPS, analysis depth, and model optimization. Typically, 1 to 3 GPUs may suffice, but exact numbers should be confirmed with onsite data.

Pilot deployments and performance testing are essential for proper hardware planning.

Inference is the process where an AI model analyzes incoming video data and generates results. This happens in real-time or near-real-time.

It is the main determinant of AI performance and latency.

Latency refers to the time between capturing an image and receiving the AI analysis results. Low latency is critical for security and automation.

Hardware, network speed, and model complexity directly impact latency.

A false positive occurs when AI triggers an incorrect alarm; a false negative is when it misses a real object. Both error types affect system reliability.

Achieving low error rates requires application-specific tuning and calibration.

Calibration adjusts cameras and AI systems for accurate and consistent data. This includes optimizing lighting conditions, camera angles, and lens settings.

Performing calibration during pilot testing improves data quality.

Behavior analysis uses AI to examine movements of people or objects captured by cameras, detecting predefined risky or unusual behaviors. It’s used in security and quality control.

The process involves extracting movement data, classifying it, and generating alarms.

Under GDPR compliance, AI video analytics follow an anonymized approach that avoids processing identity information for most modules. Facial recognition is excluded, focusing on anonymous data use.

This ensures legal compliance while maintaining effective security and analytics.

On-premise means that all video analytics and AI processing occur within the organization’s own data centers or local servers. This enhances data security and reduces dependence on internet connectivity.

This approach is preferred in industrial and security applications for low latency and maximum control.