EmotionAnalysisSystem

Identify and analyze the facial expressions of left-behind children to generate reports on their emotional trends.

Quick setup

Option 1: Using Anaconda (recommended for GPU support)

Create and activate the conda environment:

conda env create -f environment.yml
conda activate emotion-analysis

Option 2: Using venv and pip

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

FER2013 Emotion Recognition — Training scripts

This folder contains a minimal PyTorch training pipeline for the FER2013 dataset using a ResNet18 backbone.

Files added

• requirements.txt: Python dependencies (pip)
• environment.yml: Anaconda environment specification
• model.py: model factory and checkpoint loader
• data.py: dataloaders using torchvision.datasets.ImageFolder
• train.py: training script with checkpointing and fine-tune support
• face_detection.py: RTSP stream face detection and cropping
• face_detection.py: OpenCV Haar Cascade face detection (image/video/webcam/RTSP)

Training example (default expects dataset/FER2013/archive/train and dataset/FER2013/archive/test):

python train.py --data-dir dataset/FER2013/archive --epochs 20 --batch-size 64 --output checkpoints

Fine-tuning example (load a checkpoint's model weights only):

python train.py --data-dir dataset/FER2013/archive --pretrained-weights checkpoints/best.pth --epochs 10

Resume full training (loads optimizer and scheduler states):

python train.py --data-dir dataset/FER2013/archive --resume checkpoints/checkpoint_epoch10.pth --epochs 20

Notes

• The scripts use torchvision.datasets.ImageFolder, so ensure the dataset/FER2013/archive/train and dataset/FER2013/archive/test folders contain subfolders per class (e.g. happy, sad).
• Adjust --img-size if you prefer other input resolutions.
• Check --device to force cpu or cuda.

OpenCV Face Detection (face_detection.py)

face_detection.py provides a lightweight, CPU-friendly face detector using OpenCV's Haar Cascade. It does not require YOLO or GPU and supports image, video, webcam and RTSP sources. The script will attempt to locate a local Haar cascade and will automatically download it to models/cascades/ if not found.

Basic usage examples (PowerShell):

• Detect faces in a single image and save crops:

python face_detection.py --source Q.jpeg --output-dir faces_opencv --save-crops

• Run on a video file (detect every 5 frames, show preview):

python face_detection.py --source video.mp4 --interval 5 --save-crops --display

• Use RTSP stream (detect every N frames specified by --interval):

python face_detection.py --source "rtsp://user:pass@192.168.1.1:554/stream" --interval 300 --duration 300 --save-crops

rtsp stream

In our project, we use rtsp://admin:CUUNUZ@192.168.137.230:554/h264/ch1/main/av_stream as our Ezviz monitor's rtsp stream.

Key options:

• --cascade: Path to a Haar Cascade XML file. If omitted the script will auto-locate or download haarcascade_frontalface_default.xml.
• --scale-factor, --min-neighbors, --min-size: Tune detectMultiScale parameters to reduce false positives or detect smaller faces.
• --interval: For video/RTSP, detect every N frames to reduce CPU usage.
• --save-crops: Save detected face crops to --output-dir.
• --display: Show a preview window for video processing.

Notes:

• Haar Cascades are fast and suitable for lightweight detection or when GPU/YOLO is not available, but they are less accurate than modern deep-learning detectors. Use YOLO flow when higher accuracy is required.
• Downloaded cascade files are stored under models/cascades/ in the project for reuse.