FallGuard is an end-to-end computer vision system that detects whether a person is standing or has fallen from images or video. It performs bounding-box localization and posture classification using a custom-trained deep learning model built with TensorFlow.
Demo Video 
🔗 Streamlit App: https://fallguard-defi.streamlit.app
Note: The live demo runs on CPU-only infrastructure (Streamlit Community Cloud). Real-time inference may be slower than local GPU execution. A recorded demo video is provided above to showcase full performance.
- Real-time Detection: High-accuracy fall vs. stand classification.
- Spatial Localization: Bounding-box detection of the subject.
- Multi-Modal Input: Supports webcam capture, image uploads, and MP4 video files.
- Dynamic UI: Adjustable confidence thresholds via an interactive Streamlit dashboard.
- Backbone: MobileNetV2 (pretrained on ImageNet).
-
Dual-Output Heads: - Bounding-box regression: Predicts spatial coordinates (
$x_{min}, y_{min}, x_{max}, y_{max}$ ).- Binary classification: Predicts posture state (Fall / Stand).
- Framework: TensorFlow (tf.keras).
The model was trained on a dataset of ~7,000 annotated images with custom preprocessing and coordinate normalization to ensure stable localization and accurate predictions.
- Custom Pipeline: Data pipeline featuring bounding-box normalization.
- Optimization: Joint optimization of localization (Regression) and classification losses.
- Accuracy: Achieved ~91% classification accuracy with reliable spatial localization.
Technical Note on Training Curves:
As shown in the training_metrics.png, the Training Loss (blue line) demonstrates a smooth and stable convergence toward zero. The spike in Validation Loss (orange line) at Epoch 3 represents a temporary variance often caused by learning rate overshooting or validation batch noise. However, the model successfully recovered by Epoch 4, confirming strong generalization capabilities.
- Language: Python
- DL Framework: TensorFlow (tf.keras)
- Computer Vision: OpenCV
- Deployment: Streamlit
- Analysis: NumPy
pip install -r requirements.txt