Introduction In this project, we aim to create a position-based interaction system using Unity and InteractML. Our goal is to train a machine learning model which classify and recognize different movement states of the player based on specific positions in real-time and use this data to adjust environmental parameters accordingly, such as changing the color of environment light based on movement states. After recording several training examples we trained the classification model using InteractML, where the trained model can dynamically update environmental settings in Unity based on live position data.
Dataset
The dataset we used for this project consists of live position data recordings of an avatar in a clean lab environment. We recorded different training examples and each representing distinct states or behavior of the player. These recordings provided integer-labeled data points correspond to each state.
Feature Extraction:
• The primary feature extraction was the avatar's position coordinates (e.g., x, y, z). These features were mapped to the corresponding integer labels 0 or 1 to enable classification.
Model
We used InteractML’s Classification Model which is K-Nearest Neighbors (KNN) algorithm that classify avatar positions based on the training data we collected. The classification model takes the position data as input and outputs either 0 or 1 based on current state of the avatar. Several recording examples are needed to train and run a robust classification model.
Key Model Components:
-
Input Features: Real-time position data of the avatar.
-
Output Labels: Integer values 0 or 1.
Training and Evaluation Training Process • We manually recorded multiple training examples for each state by labeling as 0 for wrong path and 1 for the right path direction. Then we moved the avatar position differently after every input. • These recording examples were fed into the InteractML Classification Model for training which performs a binary classification. The model learned to associate position coordinates with the corresponding integer labels.
Evaluation
• After training, we tested the model that is able to classify position of the avatar and control the light based on avatar´s current state. We evaluated the performance of the model using a confusion matrix.
Results
The model efficiently classified avatar positions based on the input label. This classification allowed us to control the environment light dynamically in Unity where the color of the light turn red and green for the predicted integer value 0 and 1 respectively.
As we evaluated the accuracy of the model’s classification using confusion matrix, here is the key metrics output value from the confusion matrix:
• Accuracy: 0.90, where the model correctly classified 90% of all cases.
• Precision: 0.92, when the model predicts positive, it is correct 92% of the time.
• Recall: 0.98, the model almost correctly identifies actual positives.
• F1-Score: 0.95, suggest a good balance between precision and recall.
The performance of the model was reliable when tested with new live position data, where it is updating environmental parameters in real-time.
References
- InteractML Unity