From Vehicle Counts to Constrained Green-Time Recommendations: An Evidence-Audited YOLOv8 Traffic-Signal Prototype

The data that an object detector produces as totals of vehicles is not necessarily a reliable measure of congestion, and a list of recommended green durations is not a safe traffic controller. In this paper, an evidence-audited prototype specification is proposed, combining a vehicle detection model (YOLOv8) with recommendations for the length of green time that are expressed in an open way. The workflow can recognize the vehicles, such as motorcycles, buses and trucks, and can compare the pretrained nano and medium models against the fine-tuned nano model, and can summarize the visible demand for approach using a count-to-capacity ratio. It offers more capabilities such as persistent tracking, low motion events, short-horizon forecasts, nominal corridor offsets and an emergency priority demonstration. The problem of timing stage is translated into a projected lower and upper bounds constraint on the budget of a particular vehicle for each stage, and the total budget is to be a fixed value. This treatment must be different from the case of separately cutting the proportional duration as this may lead to the budget violation. The project material provided did not include the notebook, dataset, trained weights, execution logs, annotated outputs and traffic-simulation records. The paper thus presents a repeatable specification for the system, an analytical consistency test and a progressive validation procedure, but no evidence of accuracy or operation improvement. The quality of the detector is evaluated independently of the counting error, the reliability of the allocation, the temporal-event performance, and the network results. The design produced is ‘producible’ and ‘testable’ in that it can be tested and simulated in the future when observations are available, rather than assuming the claim given by the visual prototype is true, and it is defensible in that it is a design and not a claim.