What does an object detector actually output?
For every frame it examines, a detector produces a list. Each entry holds a bounding box, which is a rectangle given by its position and size, a class label such as person or vehicle, and a confidence score between zero and one. Nothing more. It does not know who the person is, what they intend, or whether they belong on the property. Keeping the detector's job this simple also makes it easy to swap in a better model later without rewriting the rules that depend on it.
That narrow output is a strength. Because the question is limited to 'what and where', detectors can be trained and tested rigorously, and the results plug neatly into rules. Everything else, such as 'this person crossed a fence line' or 'this truck has been parked for a while', is built on top of these boxes by separate logic, not by the detector itself. Anything involving identity, intent or authorization has to come from other systems, people, or site knowledge, not from the detection step.
How does a detector find objects in a frame?
Modern detectors are convolutional or transformer-based neural networks. The frame is resized to a fixed input size and passed through layers that extract features at several scales, so large nearby objects and small distant ones can both be found. The network proposes candidate boxes, scores each against the classes it knows, and discards overlapping duplicates through a step often called non-maximum suppression. Some systems also output a rough outline or pose estimate, but the box and class remain the common currency passed between the stages.
Single-pass designs make all of these predictions in one sweep, which keeps them fast enough for live video. Heavier designs examine candidates more carefully and can be more accurate at the cost of compute. The choice is a trade-off between speed, accuracy and the power budget of whatever hardware runs the model, which is especially relevant for equipment that runs on batteries and solar. Input resolution is a quiet trade-off: larger frames reveal small targets but cost more computing time and more electrical power.
Why is tracking needed after detection?
A detector treats each frame as a fresh start. It has no memory that the person in this frame is the same one from a moment ago. A tracker supplies that memory by matching boxes between consecutive frames, using position, predicted motion and sometimes appearance cues. Each object gets an identity that persists while it stays in view. Good trackers also predict where an object should appear next, which helps them survive a few missed detections without losing the identity.
Tracking is what makes behavior-based rules possible. Counting, direction of travel, and how long something has stayed in an area all depend on the same object being followed over time. It is also a weak point: if an object is hidden behind a pole or another person, the tracker may lose it and assign a new identity, which can reset timers or create a duplicate event. Understanding this explains why a clear, unobstructed view often improves analytics more than any change to detection sensitivity.
What makes people and vehicles hard to detect?
Size in pixels is the biggest factor. A person far from the lens occupies only a few pixels, leaving the network too little detail to separate them from clutter. Camera angle also matters, since a model sees people differently from a steep overhead view than from eye level. Partial occlusion by vehicles, fences or foliage removes the features the network relies on. Camera height and tilt shape all of this, so a position that looks fine to a human viewer may still be awkward for a model.
Lighting and image quality complete the picture. Strong backlighting turns subjects into silhouettes, motion blur smears edges, and heavy compression discards fine detail. Vehicles bring their own quirks: headlights can bloom and hide the shape behind them, and unusual vehicles such as trailers, equipment or bicycles may be classed differently from the cars the model saw most often in training. If a target must be reliably recognized, plan the camera and lens around the smallest and farthest target you care about.
How do you choose and test a detector for your site?
Start by listing what you actually need to detect, where the camera will be, and how far the targets will be from it. A parking lot, a construction yard and a fence line produce very different scenes. Ask whether the classes you care about are supported out of the box, and whether the camera placement gives targets enough pixels to be recognized reliably. Keep a simple log of misses and false alerts with timestamps, since patterns often point straight to the cause, such as one angle or one hour.
Then test on real footage from your own location, covering day, night and bad weather. Record both the misses and the false alerts, and review them with the vendor to see whether the fix is a camera adjustment, a zone change, or a threshold change. A detector that behaves well in a controlled demonstration may need real adjustment to behave well on your property. Treat the first weeks as a calibration period instead of a final verdict, and expect to refine placement, zones and sensitivity together.
Key terms
- Bounding box
- A rectangle drawn around a detected object, defined by its position and size within the frame.
- Class label
- The category a detector assigns to an object, such as person or vehicle, from the list of categories it was trained to recognize.
- Non-maximum suppression
- A cleanup step that removes overlapping duplicate boxes so each real object is reported once instead of several times.
- Occlusion
- When part or all of an object is hidden behind something else, reducing the visual features available for recognition and tracking.
- Pixel density
- How many image pixels cover a given width of the real scene. Higher density gives a detector more detail to work with.
- Re-identification
- Recognizing that an object that left view and returned is the same one, so it keeps a continuous identity.
On a deployed trailer, detection quality starts with how the cameras are placed. A 22 ft telescopic mast lifts the head above parked vehicles and clutter, which helps keep people from being hidden behind obstacles. The Gen 9 head pairs one PTZ camera facing the front with three fixed cameras covering left, right and rear, so each view can be framed for the targets that matter there.
Common questions
Can object detection tell who a person is?
No. A standard detector only reports that a person is present and where. Identifying an individual is a separate face recognition task with its own legal and privacy considerations, so confirm with counsel or the local authority before using it.
Why does the system sometimes miss a person who is clearly visible to me?
People read a scene with context and memory. A model works from pixels in each frame. If the person is small, partly hidden, or blurred, the network may score them below its threshold even though a human eye infers their presence easily.
Does detecting vehicles mean it reads license plates?
Not by itself. Vehicle detection finds and classifies the vehicle. Reading plate characters needs a camera framed and exposed for that purpose and dedicated recognition software, which is a separate option from general vehicle detection.