```{eval-rst} .. meta:: :description: Learn how the Smart Tolling pipeline uses dynamic region-of-interest inference and hybrid GPU, NPU, and CPU workload distribution to achieve high throughput on edge hardware. :keywords: smart tolling, optimization, dynamic ROI inference, hybrid workload distribution, GPU, NPU, CPU, axle counting, gvapython ``` # Optimization The system achieves high-throughput processing on Edge hardware through two key optimizations defined in `config.json`. ## Dynamic ROI Inference (Hierarchical Execution) To maximize efficiency, heavy neural networks (like Axle Counting) do not run on the full 4K frame. - **Logic:** The "Vehicle Type" model runs first to find the bounding box. - **Optimization:** The Axle model is configured with `inference-region=roi-list`, forcing it to execute *only* within the coordinates of the detected vehicle. ## Hybrid Workload Distribution The pipeline maps models to available accelerators to prevent resource contention: - **GPU:** Detection and tracking (Vehicle Detection, LPR Detection, Axle Counting). - **NPU:** Classification (Vehicle Color, License Plate OCR). - **CPU:** Post-processing adapters (`gvapython`). ## Learn More - [Perception Layer](./perception-layer.md) - [Analytics Pipeline](./analytics-pipeline.md) - [Support and Troubleshooting](../troubleshooting.md)