Model Training

The system uses three models working together. This guide explains how each model is configured, trained, and saved.

---

1. YOLOv11 Knife Detector

To identify knives and weapons in video frames with high precision, we fine-tune a pre-trained YOLOv11 network.

Dataset Preparation

• The training dataset is hosted and prepared on Roboflow. It is labeled with bounding boxes around knives and similar stabbing weapons.
• The dataset should be exported in the YOLO format, generating a data.yaml configuration pointing to image splits.

Training Command

Run the following CLI command to fine-tune the model:

yolo detect train data="knife-1/data.yaml" model=yolo11n.pt epochs=181 imgsz=640

• Model: Starts from the lightweight, high-speed yolo11n.pt base weights.
• Epochs: Trained for 181 epochs.
• Image Size: Resized to 640x640 pixels.
• Result: Saves the best weights under runs/detect/train/weights/best.pt, which should be placed in models/knife/run2/weights/best.pt.

---

2. YOLOv11 Pose Estimator

The pose estimation model tracks coordinates of human body joints. You can either use pre-trained weights (yolo11n-pose.pt trained on the COCO dataset) or fine-tune them.

To train the model on a custom pose dataset:

yolo pose train data=coco-pose.yaml model=yolo11n-pose.pt epochs=100 imgsz=640

---

3. Bidirectional LSTM Classifier

The LSTM network is responsible for analyzing the keypoint sequences extracted by preprocess_videos.py and determining if they represent a violent stabbing action.

Script Config

The training parameters are configured at the top of src/train_lstm.py:

• MAX_FRAMES = 150: The target length for keypoint sequences. Shorter sequences are padded; longer ones are truncated.
• MASK_VALUE = -999.0: Floating-point placeholder value used to represent empty frames.
• BATCH_SIZE = 32
• EPOCHS = 100

Model Architecture

The Keras sequential architecture is structured as follows:

Layer	Type	Specifications	Description
1	Masking	mask_value=-999.0	Skips padded frames to focus computation on active movements.
2	Bidirectional LSTM	64 units, returns sequences	Evaluates coordinate movements forwards and backwards in time.
3	Dropout	Rate: 0.3	Mitigates overfitting.
4	Bidirectional LSTM	32 units	Refines temporal context representations.
5	Dropout	Rate: 0.3	Mitigates overfitting.
6	Batch Normalization	Default	Stabilizes gradients and accelerates training.
7	Dense	32 units, Activation: ReLU	Dense classifier layer.
8	Dropout	Rate: 0.2	Mitigates overfitting.
9	Dense (Output)	1 unit, Activation: Sigmoid	Outputs probability value from 0.0 (Safe) to 1.0 (Violent).

Data Preparation & Training Flow

1. Load npz Data: Automatically scans datasets/lstm_dataset/ for .npz files, loading input arrays (X) and targets (y).
2. Padding: Pads sequence inputs using Keras pad_sequences(..., padding='post', value=-999.0).
3. Data Split: Splits loaded samples into 80% Training and 20% Validation sets, maintaining class distributions.
4. Class Weighting: Since databases may have unequal counts of violent versus non-violent samples, the script computes class weights to prevent model bias:

   class_weights = compute_class_weight('balanced', classes=np.unique(y_train), y=y_train)

5. Optimization: Compiled using the Adam optimizer and binary_crossentropy loss.

Training Callbacks

• Model Checkpoint: Automatically monitors val_loss and saves the best iteration to models/lstm_violence_detector_tmp.keras.
• Early Stopping: Prevents training when validation loss stops improving after 10 epochs.
• Reduce Learning Rate on Plateau: Multiplies the learning rate by 0.5 if validation loss stagnates for 3 consecutive epochs.

---

Running the Training Script

Execute the training script from the src/ directory:

python train_lstm.py

After training completes, retrieve the output file at models/lstm_violence_detector_tmp.keras and rename it appropriately for integration with the main inference loop.