Master Computer Vision with Deep learning, OpenCV4 & Python

Learn practical image manipulation in OpenCV with NumPy and Python, including creating zeros and ones, cropping, masking, and visualizing edits with montages.

Explore interactive computer vision with OpenCV mouse events to scroll through directory images using a custom image viewer, guided by glob loading of pngs and on-screen navigation arrows.

Extract and display image contours using canny edges, then analyze contour hierarchy and features such as centroids, approximations, convex hulls, and orientation.

Explore the theory behind image features and keypoints, including corners and descriptors, with Harris corner detector insights, invariance, and the relationship between SSD, covariance, and eigenvalues.

Extract and visualize image keypoints using the Harris corner detector and the Cheetham detector (good features to track) in a grayscale pipeline, preparing data for subsequent descriptors.

Explore feature extraction and visualization of image features in OpenCV, extracting keypoints and descriptors with SIFT and ORB, and compare Harris corner detection to these methods.

Learn object detection by localizing and classifying objects in images or videos, using techniques from Haar cascades to YOLO, with applications in self-driving cars and security.

Compare Haarcascade limitations to deep learning's automatic feature selection and study YOLO as a one-stage detector using a grid, bounding boxes, IOU, NMS, and anchor boxes for multiple object detection.

Learn to implement YOLO object detection in OpenCV using the dnn module by loading the model config, weights, and coco names, then detect and display objects.

Train a custom object detector by transfer learning with YOLO v7 tiny on a soccer dataset using Darknet in Colab, including dataset preparation, config tweaks, and evaluation.

Explore the five major stages and three minor stages of the Deepsort algorithm, from pre-processing and detection with non-maximum suppression to Kalman filter estimation, association, and track lifecycle updates.

Build a live surveillance system using deepsort to count people, display trajectories, and focus on suspicious individuals in live CCTV footage, with on-screen counts in the top right.

Display trajectories by extracting centroids from bounding boxes, storing current and previous positions per track in a queue, and drawing trajectories on a fixed mask before overlay.

Focus on a suspicious individual by selecting a point in the graphical user interface to define a bounding box and track id, then save data to a csv with timestamps.

Align faces using 68 landmarks for consistent orientation, then encode each face into 128-dim embeddings with a single call to face encodings, enabling recognition in milliseconds.