Filter Detections Based on Class Input List

[cczarnuch]

We need to filter the detections coming from YOLOv5 into Deep SORT by the following classes:

• 1: Person
• 2: Bicycle
• 3: car
• 4: motorcycle
• 6: bus
• 8: truck

The numbers correspond to the COCO class enumerations as given in https://github.com/McMasterAI/BNA-traffic-mapper/blob/update-readme/yolov5/data/coco.yaml. I'm not sure if this is useful information.

[icecap360]

When the camera gets a clear view of the bench (no person obstructing the bench) yolov5 starts thinking the bench is a car (yolov5s.pt with image size 640). Here are some things I tried:

• yolo5l, image size 640 -> no car detected (good)
• yolo5l, image size 416 -> no car detected (good)
• yolo5m, image size 640 -> no car detected (good)
• yolo5m, image size 416 -> no car detected (good)
• yolo5s, image size 1024-> no car detected (good)
• yolo5s, image size 800 -> no car detected (good)
• yolo5s, image size 640-> car detected (bad) CONCLUSION: Either move to yolov5m or increase the image size, the optimal parameter values require testing.

Whenever we start physically testing the computer vision system, we should also try out various yolov5 models and image sizes, and determine the best one based on metrics and runtime.