Choosing the right camera for AI video analytics means selecting hardware that gives the AI model enough visual information to make accurate detections — the right resolution, frame rate, low-light performance, lens angle, and compression format for the specific deployment environment. A bad camera choice cannot be compensated for by a good analytics platform. This guide reframes camera selection as an AI inference problem and gives integrators a practical specification checklist.
AI inference quality is bounded by input quality. The detection pipeline processes video frames — and if those frames lack sufficient detail, contrast, or clarity, the model cannot extract the information it needs to make confident detections.
A 720p camera with poor low-light performance feeding a state-of-the-art AI model will produce worse results than a well-specified 1080p camera feeding a good model. Camera selection is not a separate decision from analytics platform selection — it is part of the same system design.
The question is not 'does this camera show me what I need to see' — it is 'does this camera give the AI enough information to make accurate detections.' That reframe changes the specification priorities significantly.
Minimum 1080p (2MP) for reliable person detection at standard distances up to 15 metres. 4MP or higher for large open areas, licence plate reading at range, or identification requirements where facial detail matters. Higher resolution increases bandwidth and storage requirements proportionally — balance against network capacity using a bandwidth calculator.
15fps is the practical minimum for AI detection — below this, fast-moving objects may be missed between frames. 25fps is recommended for tracking moving objects, vehicles at speed, or scenarios where smooth tracking trails are operationally important. Higher frame rates increase bandwidth significantly — only specify above 25fps if the use case explicitly requires it.
The most commonly underspecified attribute in camera procurement. True day/night cameras with quality IR illumination are essential for outdoor or low-light deployments. Check the IR range against expected detection distance — a camera with 30m IR range deployed to cover 50m of fence line will have a 20m gap of degraded detection. Colour night vision cameras provide better object classification than greyscale IR but require more ambient light to function.
Wide-angle lenses cover more area but reduce object size at distance. Narrow lenses provide more detail at range but require more cameras for full coverage. For AI analytics, the rule of thumb is: the target object should occupy at least 10% of frame height for reliable detection. Calculate this from camera height, distance to target, and lens angle before specifying — a person at 30m from a wide-angle 2.8mm lens on a 1080p camera may be too small for confident detection.
Specify H.265 for all new installations — it saves approximately 40–50% bandwidth and storage compared to H.264 at equivalent quality. Verify that the analytics platform supports H.265 decoding before finalising the camera specification.
Verify ONVIF Profile S or T compliance to ensure compatibility with the analytics platform. Ask for the specific profile version, not just 'ONVIF compatible' — compatibility varies across profile versions.
Placement affects detection reliability as much as camera specification. The best camera placed in the wrong position will underperform a modest camera placed correctly.
Camera height: 3–4 metres for indoor person detection; higher mounting for vehicle monitoring or large-area coverage. Avoid mounting above 6 metres for person detection — the overhead angle compresses the human silhouette and reduces detection confidence.
Angle to subject: Perpendicular to the direction of travel gives the best detection accuracy. A camera looking down a corridor (parallel to movement) detects people as they approach head-on — which is the most difficult angle for silhouette-based detection. A camera mounted to the side of a corridor (perpendicular) captures the full body profile.
Avoid backlit scenes: Cameras pointing toward windows, doorways to bright exteriors, or strong light sources generate silhouettes that the AI struggles to classify. If the camera must face a light source, specify a model with Wide Dynamic Range (WDR) — this captures detail in both bright and dark areas of the same frame.
Overlap between cameras: Plan for 10–15% overlap between adjacent camera views to ensure no detection gaps at boundaries. This overlap also supports cross-camera tracking.
PTZ vs fixed: Fixed cameras provide consistent, always-on AI coverage of a defined area. PTZ cameras cover larger areas but require AI-triggered control to maintain effective coverage — without AI, a PTZ sitting on a preset is an expensive fixed camera with more maintenance requirements.
Organisations adding AI analytics to an existing camera estate do not always need to replace every camera. The compatibility checklist for existing cameras:
When existing cameras fall below the minimum specification, prioritise replacement of cameras covering high-risk areas first — perimeter cameras, access control points, and high-value asset areas. Lower-priority cameras can remain on the existing specification with the understanding that detection accuracy will be lower in those views.
SafetyScope publishes minimum and recommended camera specifications for every supported detection type. The platform accepts streams from any ONVIF-compliant IP camera via RTSP, with no proprietary hardware lock-in. Before deployment, SafetyScope's integration team verifies camera compatibility and provides specific recommendations for resolution, frame rate, and compression settings based on the deployment environment and detection requirements.
For organisations with mixed camera estates, the platform adapts its processing pipeline per camera — applying appropriate confidence thresholds and detection parameters based on each camera's capabilities.
Published: 2026-02-02 · Updated: 2026-04-02