Region-of-interest based video encoding

A video processing unit comprising:

• an artificial intelligence accelerator including:
• a non-object-based region-of-interest detection neural network configured to receive a video frame and generate a plurality of candidate non-object-based region-of-interest blocks for the video frame, wherein the plurality of candidate non-object-based region-of-interest blocks is generated by excluding identification of an associated object type of an object within the video frame;
• a threshold selection module configured to receive the plurality of candidate non-object-based region-of-interest blocks and identify a plurality of selected non-object-based region-of-interest blocks based on a predetermined threshold;
• a region-of-interest map generator configured to receive the plurality of selected non-object-based region-of-interest blocks and generate a region-of-interest map;
• and
• an object-based region-of-interest detection neural network configured to receive the video frame and generate a plurality of candidate object-based region-of-interest blocks for the video frame;
• and
• a central processing unit including:
• a sorting and non-maximum suppression (NMS) module configured to receive the plurality of candidate object-based region-of-interest blocks, sort the plurality of candidate object-based region-of-interest blocks for respective object types based on associated probabilities, and combine multiple overlapping object-based region-of-interest blocks to determine one or more region-of-interest bounding boxes for the respective object types.