Depthai: Harris Filtering Support

[Luxonis-Brandon]

Start with the why:

Corner detection (Harris corner filtering) is often useful in a variety of CV/AI pipelining usecases and DepthAI has hardware which allows hardware acceleration of this CV function.

Move to the how:

Implement Harris Filtering as a node in the Gen2 pipeline builder (https://github.com/luxonis/depthai/issues/136).

Move to the what:

Support Harris Filtering as a Gen2 node.

[eric-schleicher]

+1 for this getting prioritized sooner than later.

[szabi-luxonis]

This is now possible on https://github.com/luxonis/depthai-python/pull/317
Either SHI_THOMASI or HARRIS can be selected for feature detection.

featureTracker->initialConfig.setAlgorithmType(dai::FeatureTrackerConfigData::AlgorithmType::CORNER_DETECTION);
featureTracker->initialConfig.setCornerDetector(dai::FeatureTrackerConfigData::CornerDetector::HARRIS);

[asa]

Testing this with dai::FeatureTrackerConfigData::CornerDetector::SHI_THOMASI doesn't succeed. The first few frames of tracks show up when printing, but then no new tracks appear. Sometimes it starts at the beginning for a few frames, then ends for a while, then starts back up after 10 seconds or so. Inconsistent.

[asa]

After working with it more it appears there is a temporal filter on the quantity that are on screen at any one time and they are slowly added, one at a time till tracking a few hundred on an 800_P frame in my environment. They do look cleaner than the harris corners. Nice!

[asa]

Well, that might have just been because I was running the Left tracker with SHI_THOMASI and the right one with HARRIS for comparison. Setting them both to SHI_THOMASI seems to solve the issue. Sorry for the noise!