Bryan Bednarski - April 8, 2021

This repo serves as a jumping-off point for system integration of semantic instance segmentation and multi-object tracking in real-time applications. The initial goal of this work was the explore possible optimizations in semantic segmentation through tried-and-true methods of multi-object tracking (DeepSort). However, I found empirallcy that the region proposal network built into MaskTrackRCNN (using non-max supression) performs exceedingly well in proposing regions of interest (ROIs) to the rest of the network without clues from object tracking. Regardless, this system still has value in it's added object tracking functionality (kalman filter) and can hopefully be built into a physical platform eventually with a complete system pipeline at a later time.

Code on Github

Block diagram of the original investigation:

Next Post Previous Post