Object Detector Model

DISCLAIMER The object detector model project will have a heavy research emphasis

Context: During the URC competition we must be able to detect two types of objects: mallet and water bottle. These objects will exist at two distinct waypoints during the autonomous mission. For a full context on the auton mission please read the URC Rules, specifically 1.f Autonomous Navigation Mission.

Problem: The majority of the R&D went into the object detector's execution/inference environment, leaving the model left somewhat rudimentary. The current model is a YOLOv8 small with fine tuning using our own mallet and water bottle dataset and RoboFlow.

Solution: First Step:

• Get the model to work with YOLOv8n it is a lighter weight model and a great starting point.
• Perform fine tuning using RoboFlow

Future Research: Use a smaller, lighter 🪶 model using one of:

• FIND A MODEL THAT ACCEPTS BGRA/Non-Blob format: we are currently spending more time doing image conversions than actually processing the data.
• Tensor Flow:
  • Instructions to do
  • Model Bank
  • The smallest one I found was 9.4 MB
  • Convert .tflite to onnx
• YoloV8:
  • YoloV8n has a size of 6.4 MB
• Training Custom Model (Using YOLO)
  • Process

---

Interface (subject to change): Use the existing TensorRT framework to execute the model and then publish the objects into the tf tree. Here are some of the tf tags:

bottle
mallet
immediateMallet
immediateBottle