Feature Extraction in Robotics Vision

What is it?

Feature extraction transforms raw sensor data into compact descriptors that capture the essence of what's being observed — typically corners, edges, or learned embeddings. Without feature extraction, robots would have to process every pixel from scratch every frame.

How it works

Classical feature extractors like SIFT, SURF, ORB, and FAST detect interest points in an image and produce a fixed-length descriptor for each — invariant to scale, rotation, and (somewhat) illumination. Modern deep-learning features are produced by convolutional networks (e.g., ResNet features) or self-supervised transformers (e.g., DINO).

Real-world example

Visual SLAM systems like ORB-SLAM rely entirely on ORB features. Visual place recognition for autonomous vehicles uses learned features. Boston Dynamics Spot's localisation uses both classical and learned features.

Why it matters for robotics

Feature extraction underpins every perception system. Choosing the right features (classical vs learned, dense vs sparse) is a fundamental design trade-off that every robotics engineer faces.

See also

• Computer Vision
• Object Detection
• Slam