Gaits prediction

[ModusMorris]

feat: Add multi-label gait recognition alongside step detection

• Introduce 7-output CNN (1 step + 6 gait types) using BCE for multi-task learning
• Shift dataset logic to include a 6D gait vector per window
• Adjust loss function to handle weighted step and gait losses
• Modify prediction script to compute argmax gait prediction and visualize it in the final plot

[ModusMorris]

Possible Reasons for Inaccurate Step and Gait Recognition

Data Quality and Imbalance

If some gait types (e.g., “frei_mitschwingend”) dominate the training set, the model leans toward predicting them more often.
Errors in step or gait labels (e.g., mislabeled peaks) also mislead the network.
Window Size and Labeling Offset

Large windows (e.g., 128 samples) dilute exact step information, making peak localization harder.
Shifting labels by half the window size can cause systematic timing offsets between true peaks and the model’s predictions.
Multiple Gait Labels

Having multiple gait labels active at once complicates classification (it’s harder than single-class).
Overlapping or very similar gaits (e.g., “langsames_gehen” vs. “normales_gehen”) are difficult to distinguish from raw sensor data.
Model Hyperparameters

Incorrect loss weighting (steps vs. gaits) can skew training.
Too few epochs, high dropout, or suboptimal learning rate reduce the model’s accuracy.
Mismatch Between Training and Inference

Using a different window size or normalization for prediction than for training leads to inconsistent dimensions and poor results.