The Vehicles and Pedestrians Detection application demonstrates ModelPack trained to detect various vehicles and pedestrians as found in the Berkeley Diverse Driving dataset. The application is run on an NXP i.MX 8M Plus EVK with an OV5640 camera running at 1080p and displaying results to an HDMI display running at 1080p. The model is run at 416x416 and maintains a consistent 60 FPS.
Note: the camera driver is limited to 30 FPS but for the demo we double sample each frame to demonstrate that a 60 FPS pipeline is achievable.
The application includes a metrics overlay showing the camera, display, and inference timing details. We can see the model has excellent performance on the NPU achieving sustained 60 FPS. The core inference time is around 12ms while the input and output processing is about 4ms.