Fall Detection Capstone Project

1. Introduction (2 trang) Khoa’s survey (T7 25/11)

1.1 Motivation ✅

1.3 Scope and Objectives of the Project ✅

2. Technological Foundations (T7 25/11)

2.1 Deep learning Basics @Nam Anh Bui ✅

2.1.1 Deep Learning Overview ( Classifier problem, detection … ) ✅

2.1.1.1 Neural Networks ✅

2.1.1.2 Convolutional Neural Networks (CNNs) ✅

• Convolutional Operation ✅
• Convolutional Neural Network ✅
• Pooling Layer ✅
• Feature Map ✅
• Object Detector Problem [ ]

2.1.1.3 Graph Convolutional Network @Phú Xuân ✅

2.2 Human Pose Estimation (T5, 30/11)

2.2.1 Data Preprocessing Khoa’s survey ✅

• Image Resizing and Normalization: Images are often resized to a specific input size suitable for the pose detection model. Normalization may be applied to standardize pixel values.
• Data Augmentation: Techniques such as rotation, scaling, and flipping are used to augment the training dataset, improving the model's generalization.

2.2.2 Pose Detection Model: Khoa’s survey ✅

• Convolutional Neural Network (CNN): Deep learning models, especially CNNs, are widely used for pose detection. The architecture may include multiple convolutional layers to extract hierarchical features.
• Architectures for Pose Detection: Popular architectures include Hourglass Networks, OpenPose, PoseNet, and HRNet, each designed to capture spatial relationships and hierarchical features relevant to human pose.

2.2.3 Heatmap Regression: Nam Anh survey

• Joint Heatmaps: For 2D pose estimation, joint positions are often represented as heatmaps. Each heatmap corresponds to the likelihood of a joint's presence in a specific location.

2.2.4 Part Affinity Fields (PAFs): ✅

• PAFs for Limb Connections: PAFs connect pairs of joints, indicating the likelihood of a limb existing between them. They help refine joint positions and capture the spatial relationships between joints.

2.2.5 Post-Processing: Nam Anh survey

• Non-Maximum Suppression (NMS): To remove redundant and closely located joint predictions, NMS is applied. It retains only the most confident predictions for each joint.

2.2.6 Skeleton Representation: Nam Anh survey

• Connecting Joints: After detecting joints, a skeleton representation is formed by connecting adjacent joints with lines, representing the limbs.

2.2.7 Output Format: Khoa

• Joint Coordinates: The final output includes the estimated coordinates (x, y) of each joint in the image. In 3D pose estimation, depth information (z) is also included.

2.2.8 Multi-Person Pose Detection: @Phú Xuân ✅

• Person Detection: In multi-person pose detection, a person detection model is often used to identify and locate individuals in the image before estimating their poses individually.

2.2.9 Attention Mechanisms (Optional): @Phú Xuân ✅

• Attention for Body Parts: Some models may incorporate attention mechanisms to focus on specific body parts or regions of interest during the pose detection process.

2.2.10 Dataset @Phú Xuân ✅

• Annotated Pose Dataset: Training and evaluating pose detection models require datasets with annotated images containing ground truth pose information.

2.2.11 Human Fall Detection Khoa’s survey

• 1.4 Traditional Approaches (smart watch, sensor, ...)
• 1.5 Proposed Solution

2.2.12 Application @Phú Xuân ✅

• Protocols
• UI / UX
• Web basic

3. Methodologies (CN tuan sau)

3.1 Frameworks

• 1. ST-GCN (@Phú Xuân )
• 2. Alphapose Khoa’s survey + @Nam Anh Bui

3.2 Hardware @Phú Xuân

• 1. Jetson Xavier

3.3 Application

• 1. React Native @khoa Truong
• 2. Django @nam Anh Bui
• 3. PostgreSQL @nam Anh Bui
• 4. Cloudinary @Phú Xuân
• 5. Docker @Phú Xuân

4. System Architecture

• RSTP: Tim hieu protocol, cach hoat dong (co thong qua gateway, …) @Phú Xuân
• Cloudinary + CDN?: Setup local…
• Notification: Nam Anh survey
• Api gateway: @khoa Truong
• How device connect to wifi through mobile app.

5. Evaluation

6. Planning @khoa

www.notion.so