Histogram of Oriented Gradients (HOG)
The Histogram of Oriented Gradients (HOG) is a computer vision technique that describes an image by analyzing the local intensity gradients and their orientations. The image is divided into small cells, and each cell generates a histogram of gradient directions, which together form a descriptor useful for object detection.
Background
Introduced by Dalal and Triggs (2005) for pedestrian detection, HOG became a widely used descriptor in vision tasks, often combined with Support Vector Machines (SVMs). While deep learning has largely replaced HOG in modern pipelines, it remains valuable in lightweight applications requiring speed and low computational cost.
Applications
- Pedestrian detection in automotive safety systems.
- Video surveillance for silhouette recognition.
- Document analysis for detecting shapes and characters.
Strengths and challenges
- ✅ Computationally efficient and interpretable.
- ✅ Does not require large labeled datasets.
- ❌ Outperformed by CNN-based methods on complex tasks.
- ❌ Sensitive to lighting changes and scale variations.
HOG was revolutionary when it appeared because it offered a way to capture shape and edge information without relying on pixel values alone. By encoding how gradients are distributed, it effectively describes the contour and silhouette of objects, which is especially useful for tasks like pedestrian detection where outlines matter more than textures.
The technique became a standard baseline in computer vision research during the late 2000s. It was often combined with sliding window approaches and linear SVMs to detect objects in images. Although convolutional neural networks later surpassed HOG in accuracy, HOG descriptors are still popular in embedded systems and robotics, where computational budgets are limited and interpretability is valued.
An additional strength is that HOG is rotation- and illumination-sensitive in a predictable way, which can sometimes be beneficial. For instance, in controlled environments (e.g., industrial inspection), these sensitivities act like a natural filter, making it easier to detect subtle shape differences.
📚 Further Reading
- Dalal, N., Triggs, B. (2005). Histograms of Oriented Gradients for Human Detection.
