An illustrative example of the AM62A's capabilities is a security camera application. Video frames are captured through the MIPI CSI-2 port using a raw output camera. The processor sends configuration parameters for gain and white balancing back through the same interface. The raw output frame is denoised and de-mosaiced on AM62A's VPAC-3L image sensor processor (ISP), producing usable RGB images. If a wide-angle lens is used, then the lens distortion correction (LDC) accelerator within the ISP will transform the image to reduce warping. The VPAC-3L ISP can further preprocess the image to match the input specification of a machine learning model, such as an object detection network to recognize humans, vehicles, and animals. This ML model runs on the 2 TOPS deep learning acceleration hardware within the processor. The result of the model may be used to trigger an alarm or uplink an H.265-encoded clip of the event over the network. Figure 3-1 depicts this example at a high level, and Figure 3-2 shows the block diagram for example values for resolution, framerate, and pixel-format.

For smart camera applications like security or surveillance, object/keypoint detection deep learning models are most relevant as they may recognize and localize multiple things, such as humans and vehicles, within an image.