Real-Time & Fast Object Detection and Display with YoloV5

Uses three threads and three queue's for fast model inference without spending time waiting for screenshots or images with labels to be rendered to an opencv window.

How it works

Screenshotter Thread:

• Do screenshot ->
• Resize screenshot ->
• Add to requires-inference queue ->
• Repeat ->

Inference Thread:

• Get image from requires-inference queue ->
• Feed image to model ->
• Put image and image results onto draw-objects queue ->
• Repeat ->

Draw Object Thread:

• Get Image from draw-objects queue ->
• Process result boxes and labels and draw them on image ->
• Add to prepared-screenshot queue ->
• Repeat ->

Main Thread:

• Wait for prepared screenshots from prepared-screenshot queue ->
• Show screenshot ->
• Repeat ->

Honestly idk what to call this section

My Laptop with an 80 watt 3060 mobile was able to achieve 38 fps with the following global variable settings, using python 3.9.6.

# Change to 'cpu' for cpu inference.
DEVICE = "cuda"  
# Change to lower number like ~10-30 if you want for slower fps and less cpu / gpu usage.
DESIRED_FPS = 100  
# Screenshot capture area on the screen (single monitor)
DESIRED_SCREEN = {"top": 0, "left": 0, "width": 1920, "height": 1080} 
# Size after resizing from DESIRED_SCREEN
DESIRED_INFERENCE_SIZE = (960, 540)
# Model name to pull from torch hub. Models are described / documented on this page: https://github.com/ultralytics/yolov5
YOLOV5_MODEL_NAME = "yolov5m6"

References

Fast Object Detection: yolov5