A city authority relied on periodic manual counts to plan signal timing and road changes. The proposed system derives continuous vehicle counts, classifications and turning movements from existing junction cameras while keeping all processing anonymous.
The authority planned signal timing and junction improvements using traffic counts collected by survey teams a few days each year. Those counts were expensive, quickly out of date and blind to seasonal, weather and event-driven variation.
Existing junction cameras were used only for live viewing by the control room. Nobody had the capacity to extract structured data from them, and the authority did not want to install new sensors at every site.
Planners wanted vehicle counts by class and turning movement, queue length estimates and detection of stopped vehicles or wrong-way driving, all available as continuous data rather than snapshot surveys, and without identifying vehicles or people.
Junction cameras were installed for human viewing, not analytics: wide angles, oblique viewpoints and partial occlusion by signals and street furniture made consistent detection and tracking harder than a purpose-mounted sensor.
Turning-movement counts require tracking each vehicle from an entry approach to an exit approach across the whole junction, including through dense queues where vehicles occlude one another.
Classification between cars, vans, trucks, buses, motorcycles and bicycles had to remain reliable at night, in rain and under low sun, conditions in which the survey teams had traditionally struggled as well.
The authority required that no plate reading or face identification take place, and that only aggregated counts and anonymised event records leave the junction.
NVIDIA Jetson devices in roadside cabinets consume the existing camera feeds and run vehicle detection, classification and tracking locally.
Each camera view is calibrated once to map image positions to junction approaches and lanes, enabling turning-movement assignment and queue-length estimation.
Counts are aggregated into fixed intervals per approach and class; stopped-vehicle and wrong-way events are emitted individually with a short anonymised clip.
Aggregated data flows to a cloud-hosted PostgreSQL store exposed through a FastAPI service for planners, with dashboards for the control room.
Reviewed camera coverage at candidate junctions and selected sites where existing views could support reliable tracking, recommending camera adjustments where they could not.
Fine-tuned detection and classification models on footage from the selected junctions, with emphasis on night, rain and low-sun conditions.
Validated automated counts against manual survey counts on the same days before the authority accepted the data for planning use.
Deployed edge units in roadside cabinets with remote update and health monitoring, so maintenance did not require site visits for software changes.
Delivered planner dashboards and a data export interface aligned with the formats the authority already used for traffic modelling.
Traffic counts and turning movements are available continuously rather than from periodic surveys.
Signal timing reviews can draw on data covering different seasons, weather and events.
Stopped-vehicle and wrong-way incidents are surfaced to the control room as they occur.
All processing is anonymous; only aggregated counts and anonymised event clips leave the junction.
People counting, traffic analytics and public-space monitoring that give city operators a live, measurable view of how spaces are used.
Vehicle detection, classification, tracking and identifier recognition for roads, sites, fleets and transport hubs.
Real-time video understanding, event detection, behavioral analysis and multi-camera analytics.
We can walk you through how this approach would translate to your environment and what it would take to pilot it.