The same model, with the same software version, produces reliable alarms on one camera and is useless on the camera next to it. On site, the software is the first suspect; often, the problem is how the camera sees the scene. Camera angle for analytics, the target’s pixel size, occlusion, motion blur and frame rate directly determine the model’s decision. Below we unpack these effects with numerical examples, and give geometry requirements by type of analytics and the checklist we use to check an existing camera’s suitability for analytics.
How the model sees the camera
A human viewer interprets the frame as a whole and fills in missing information from context. An object detection model doesn’t; it works with a fixed-size input tensor. The typical flow is:
- The stream from the camera (often not the main stream but a lower-resolution substream) is decoded.
- The frame is scaled down to the model’s input size; in most implementations, the aspect ratio is preserved and the remaining area is filled with padding (letterbox).
- The model produces a class and confidence score for each candidate region.
- The tracker combines detections across consecutive frames; the rules engine evaluates zone, line and duration conditions on the track.
Two consequences follow from this chain. First, the resolution the model “sees” isn’t the resolution on the camera’s label but the resolution left after scaling down. Second, even if detection is correct on a single frame, the rule doesn’t trigger if the tracker loses the target across consecutive frames. Camera geometry affects both stages.
The model doesn’t read the label.
The rules of camera placement for a human viewer (face visibility, mounting height, blind spots) are in the camera mounting height and angle article. Here we look at the same scene through machine vision.
Object pixel size
For the model to detect reliably, the target must be above a certain pixel height in the scaled-down image. This lower limit varies by model and manufacturer; manufacturers usually give the minimum target size in pixels or as a ratio of image height. In the design, take this value from the documentation of the product you’re using, not from another product or forum information.
Worked example: consider a camera 1920 pixels wide with a horizontal field of view of 90°.
- At 20 m, the scene width is 2 × 20 × tan(45°) = 40 m; pixel density is 1920 / 40 = 48 pixels/meter. A 1.7 m tall person is about 82 pixels high at full resolution.
- If the analytics engine scales the frame down to 640 pixels wide, density drops to 16 pixels/meter and the person’s height to about 27 pixels.
- The same camera sees a person at 10 m as about 54 pixels after scaling down.
The assumption “let’s get a 4K camera so it sees far away too” often comes to nothing because of this calculation: if the engine also scales the 4K frame down to the same input size, there’s no gain. The gain comes if the engine splits the image into tiles, crops the region of interest or works with a larger input size. The cost is processing load. For lens selection and field of view calculations, you can use the camera lens and focal length calculation guide.
Jobs needing attributes and detail (hard hat color, vest, plate characters) need more pixels than detection. The DORI classes in the IEC 62676-4:2014 table (detection 25, observation 62.5, recognition 125, identification 250 pixels/meter) are defined for a human viewer; even so, they’re a good reference for seeing the order of magnitude of the requirement.
Tilt angle and perspective
Models largely learn the viewpoints in their training data. The training data of general-purpose models consists mainly of images shot close to eye level or slightly from above. As camera tilt increases, the target’s appearance changes:
- Low tilt (close to horizontal): Long distances are covered, and perspective is strong. Nearby targets look very large and distant ones very small; targets pass in front of each other.
- Medium tilt: Often the most balanced range for general detection, virtual fences and PPE; the target’s body and head are visible together.
- Steep view (from above): Occlusion is minimal; ideal for people counting and occupancy. On the other hand, a person appears as a head-and-shoulders silhouette, general models can be weak with this view, and face and clothing attributes can’t be obtained.
Perspective also distorts distance and speed estimation. In rules that make decisions in meters, such as forklift–pedestrian proximity analytics, the closeness of two boxes in the image plane doesn’t show the real distance; ground plane calibration is needed. At low tilt, the error of this calibration grows quickly with distance. Distance without calibration is a guess.
Occlusion and scene complexity
Occlusion is part of the target being covered by another object or person. The detection score drops, the tracker loses the target, and when it finds it again, assigns a new track number. In counting analytics, that means a double count; in intrusion analytics, the rule timer resets and the alarm never happens.
Typical site conditions that increase occlusion are rack aisles, parked vehicles, columns, crowded passages and vegetation in front of the camera. A complex background is a separate headache: reflective surfaces, screens, human figures on posters and moving machine parts produce false detections. Some of these are managed on the rule side with an exclusion zone. But an exclusion zone is also a blind area; don’t forget that when drawing it.
Frame rate and blur
Motion blur
A moving target leaves a trail on the image during the exposure. Blur is estimated approximately as speed × exposure time × pixels/meter. Worked example: a forklift traveling at 3 m/s, in an area with a density of 100 pixels/meter, produces about 10 pixels of blur at a 1/30 s exposure; at 1/250 s, this falls to 1.2 pixels. At night, auto exposure lengthens the shutter time to reduce noise and increases blur. This is the most common reason analytics that works by day collapses at night. We explained the balance of shutter, gain and noise in detail in the camera exposure settings article.
Analysis frame rate
The analytics engine often runs below the camera’s frame rate. What matters for the rule is how many frames the target is seen in within the region of interest. Worked example: a person walking at 1.4 m/s crosses a 3 m wide zone in about 2.1 seconds; at 5 frames/s analysis, they appear in about 10 frames. A vehicle passing at 36 km/h (10 m/s) crosses a 10 m field of view in 1 second and, at the same analysis rate, appears in only 5 frames. For line crossing and direction detection, the tracker needs to see the target consistently across several consecutive frames; for fast targets, plan the analysis frame rate and camera frame rate together.
Compression artifacts
Low bitrate and a long GOP produce blocking and loss of detail in moving scenes. The human eye tolerates this; the model doesn’t, and the detection score drops. Assess the bitrate of the stream analytics uses independently of the recording stream. What’s enough for recording may not be enough for analytics.
Requirements table by type of analytics
The table gives geometry priorities by type of analytics qualitatively. Take the numerical lower limits from the documentation of the product used and verify them with a pilot recording.
| Analytics | Preferred angle | Critical parameter | Typical risk |
|---|---|---|---|
| Perimeter intrusion / virtual fence | Medium tilt, long axis | Minimum target size at the far end | Night noise, animals |
| People counting | From above or close to vertical | Complete coverage of the passage line | Targets merging in crowds |
| PPE detection | Medium tilt, short distance | High pixel count on head and torso | Hard hat ambiguity at a top-down angle |
| Vehicle–pedestrian proximity | High mount, wide coverage | Ground calibration, frame rate | Distance error from perspective |
| License plate recognition | Low horizontal and vertical angle to the plate | Character height, short shutter | Headlight glare, blur |
| Attribute search | Close to eye level at passage points | Enough pixels on the torso, color image | Loss of color in IR mode |
The effect of angle and character size on license plate recognition is a separate topic; we covered it in the license plate recognition camera selection and accuracy article.
Checking an existing camera’s suitability
Before adding analytics to an existing enterprise camera system, assess whether the camera is suitable for that analytics using frames taken from recordings. This is the job to do before buying the license. The skeleton of the checklist we use:
- Which stream (main/sub) and which input size the analytics uses has been verified from the product documentation.
- The target’s pixel height after scaling down at the farthest point of the region of interest has been calculated and compared with the manufacturer’s minimum value.
- Camera tilt is in the range preferred by the type of analytics; for a top-down view, it’s been verified that the model is suitable for this view.
- Permanent sources of occlusion in the region of interest (racks, columns, parking areas) have been marked.
- Target blur and noise level have been checked on night recordings; no IR reflection or glare.
- The target isn’t lost in high-contrast situations such as backlight and headlights.
- The number of analysis frames in which the target appears in the region of interest has been calculated for the fastest expected movement.
- The bitrate and GOP values of the analytics stream don’t produce noticeable artifacts in moving scenes.
- The camera mounting is solid; the frame doesn’t shift with wind or vibration.
For a camera that doesn’t meet the list, the options are an angle adjustment, a lens or camera change, or an additional camera dedicated to analytics. Reassessing camera placement against these criteria can be planned with camera positioning and blind spot analysis work, and choosing the analytics engine and setting up rules together under AI video analytics deployment.
Frequently asked questions
Does a higher-resolution camera always improve analytics?
No. The analytics engine often scales the image down to the model’s input size; after scaling down, a 4K stream may not give different results from a 1080p stream. High resolution helps only if the engine processes the image by splitting it into tiles or crops the region of interest. Verify from the documentation which stream and which input size the product uses for analytics.
Does analytics work on fisheye cameras?
It can, but the model must have been trained on fisheye images, or the analytics must run on the dewarped image. A fisheye camera looking straight down from the ceiling can be suitable for people counting and zone occupancy. Because the target’s pixel size drops quickly toward the edge of the image, it’s limited for work such as attribute detection.
Can analytics be added without changing an existing camera’s angle?
On some cameras, yes. A camera that meets the pixel size, angle, occlusion and blur criteria in the checklist is used as it is. Those that don’t need an angle adjustment, a lens change or an additional camera dedicated to analytics. Make the decision by measurement on pilot recordings, not by guessing.
What shutter speed is recommended for analytics?
There’s no fixed value; it’s calculated from the target’s speed and the pixel density on the target. Blur is estimated as speed times exposure time times pixels per meter, and the aim is to keep it below a few pixels. Because a short shutter increases night noise, it should be assessed together with additional lighting.