This project is a crowd counting system that employs a deep learning approach, specifically a ResNet-18-based Convolutional Neural Network (CNN). It is designed to count the number of people in a crowd or video stream in real-time. The system uses a dataset of images with corresponding crowd count labels for training.
Objective: The main objective of this project is to count the number of people in a crowd or a video stream in real-time.
Key Components:
Dataset: The project uses a dataset of images and their corresponding crowd counts. Each image is associated with a ground-truth crowd count obtained from annotation data.
CrowdCountingDataset Class: This custom dataset class is responsible for loading images and their crowd count labels from the dataset. It also applies data transformations to the images.
CrowdCountingModel Class: This custom deep learning model is based on the ResNet-18 architecture. It is used to learn and predict crowd counts from input images.
Data Preprocessing: Images are loaded, converted to RGB format, and resized to a consistent size (224x224 pixels). These preprocessed images are then fed into the model.
Training Loop: The model is trained using the dataset to learn the relationship between the input images and the crowd counts. The Mean Squared Error (MSE) loss is used for training.
Real-time Crowd Counting: After training, the model is used to perform real-time crowd counting on video frames from either a camera feed or a video file. It preprocesses each frame, feeds it to the model, and displays the crowd count on the frame.
Saving and Loading Model Weights: The trained model weights are saved to a file (crowd_counting_model_weights.pth) so that they can be loaded and reused for future inference without retraining.
Usage:
The project can be used to count crowds in real-time from a live camera feed or a video file. The model can also be used to predict crowd counts for static images.
Note:
Ensure that you have the required dataset, such as images and their corresponding crowd count annotations, to train and test the model. You can adjust hyperparameters like batch size, learning rate, and the number of training epochs to fine-tune the model's performance. Overall, this project provides a foundation for building a crowd counting system using deep learning, which can be useful in various applications like crowd management, event planning, and more.
The Training and Test datasets can be found here : https://www.kaggle.com/datasets/tthien/shanghaitech/download?datasetVersionNumber=1