Define the survey question first
A crossing count, a turning movement and the number of people visible in a frame are different measurements. Before selecting a model, specify the event that matters: a person crossing a line, a vehicle completing a movement, or the occupancy of a defined area during a given interval.
The survey brief should also identify the location, time periods, road-user classes and conditions that the evidence needs to represent. Those choices guide the camera view, the processing method and the review effort.
Understand what the detector returns
An object detector identifies candidate objects in an image. TensorFlow's COCO-SSD documentation describes an output of bounding boxes, class names and confidence scores. These are observations to interpret; a list of boxes is not a completed transport survey.
For unique crossing counts, a system also needs to associate observations over time and apply a crossing rule. A person visible across successive frames must not be counted as a new person in every frame. Ambiguous paths need a defined treatment.
Build a reviewable evaluation
For a scoped pilot, we would prepare manually reviewed examples covering the conditions of interest. The evaluation should examine missed events, additional events and classification errors for the survey measure itself. A good overall total can conceal errors that happen to cancel each other out.
The review set should include difficult conditions visible at the site: crowded periods, partial obstructions, changes in light and road users near the edge of the view. Keep a record of which periods were usable and which were excluded, with reasons.
Deliver evidence people can inspect
A useful deliverable pairs the survey table with its time coverage, measurement rules, quality checks and selected reviewable examples. That gives the transport team a way to decide whether the evidence is suitable for its intended use. A capture log should identify the camera view and any interruptions, so the review can distinguish a processing issue from a gap in the footage.
RoadInsite's current browser prototype explores person detection in video frames. It is a starting point for discussing a survey pilot. A production survey would require the additional measurement logic, permissions and validation appropriate to the project.