Pi Img Processing Performance Improvement

[HiroIshikawa]

1. Make the performance measurable:
   • within opencv.

• Python performance implovement official reference
• Overall Python performance improvement strategies: Quara

2. Performance improvement potential solutions.
   • no use of threading (check_detection() should be substituted with sequential impl in the main loop)
   • no drawing boxes (comment out the box() in the main process)
   • implement dedicated video processing thread: according to this article
   • reducing resolution (currently 960x240 -> 960x120?)
   • reducing frame rate (currently 64 -> 32?)
   • cascade accuracy (->reduce the number of required images to take average out)
   • changing measuring algorithms to the one without taking averages (continuous algorithms)
   • C/C++ implementation for the detected objects processing unit

[HiroIshikawa]

Findings with initial analysis with "python -m cProfile my_code.py"

• waitKey() consumes. it may come with default usage of cv2 so there's nothing to do with it.
• videoCapture() and videoSet() takes time, which already implemented in a way that everything regarding these two methods are completed to be configured. Good to go.
• reading cascade (cascade = cv2.CascadeClassifier("cascade.xml")) every calls of the detect() method is inefficiency. => create cascade object once and pass its reference
• cap.read() to extract frame is consuming in the main loop. => the dedicated video capturing threads might mitigate this negative effect
• 'detectMultiScale' of 'cv2.CascadeClassifier' class one of the most time consuming functions.
• waitKey() somehow called without being explicitly included in the script.

=> Optimisation for PiCamera module and Pi specific environment is required. At this point,

• Do pass reference of CascadeClassifier once.
• Implement the dedicated video capturing stream thread
• Figure out the way to reduce the consumed time of the detectMultiScale / call
  • searching space = resolution. => reducing resolution makes it efficient
    • the detectMultiScale should have the params to configure searching region
• Figure out the way to eliminate the waitKey() call without alternative error

[HiroIshikawa]

Implemented and improved overall performance of object detection:

• dedicated video capturing stream thread.
• check_detection() thread implementation was substituted with main thread algorithm