Intelligent Video Surveillance Using Deep Learning System

A deep learning system that detects abnormal activity in surveillance video. It analyses frames in real time, flags unusual events such as fights or intrusions, and saves those frames for review.

Images are collected and labelled into classes, then resized and augmented. A convolutional neural network (often a fine-tuned MobileNet or ResNet) is trained on the training split. Accuracy, precision, recall and the confusion matrix are measured on unseen test images. The trained model is loaded in a Flask or Streamlit app where the user uploads an image or uses the webcam and sees the prediction.

Convolution layers slide small filters over the image to detect edges and textures, pooling layers reduce the size, and fully connected layers combine the features to classify the image. Training adjusts the filter weights with backpropagation.

Overfitting is when the model memorises training data and fails on new data. We used data augmentation, dropout, early stopping and a separate validation set to control it.

Capture the frame, convert to grayscale, remove noise with a Gaussian blur, segment the region of interest with thresholding or edge detection, extract features and finally classify or match them.

First a detector (Haar cascade, HOG or a CNN) finds the face. Then a model converts it into an embedding, a list of numbers that describes the face, and compares it with stored embeddings using a distance measure.

It follows a three-tier architecture: the browser shows the user interface, the server-side code handles logic and validation, and the MySQL database stores records. Requests travel over HTTP and responses are rendered as HTML or JSON.

We followed an iterative (incremental) model: requirements and a basic design first, then modules built and tested one by one, with the guide's feedback after each review. It suits a project where requirements become clearer as we build.

Unit testing of each module, integration testing of the complete flow and system testing with real inputs. Test cases with expected and actual results are listed in the report, including invalid inputs.

The current version is tested on a limited dataset or number of users. Future work includes deploying on the cloud, adding more features or languages, improving accuracy with more data and building a mobile version.