Algorithm 3. Visual Features Extraction

Input: M images from M directions in Node i, M equals to 4 in a square environment

Output: Bag-of-words vector $V_{ij}$, where j is the number of image in Node i

1: Create a visual vocabulary in an offline step based on DBoW3 with ORB features

2: Robot captures images from four directions, $I_{i1}$, $I_{i2}$, $I_{i3}$, $I_{i4}$

3: Extract ORB features from each image

4: Covert features of each image into a bag-of-words vector, $V_{i1}$, $V_{i2}$, $V_{i3}$, $V_{i4}$

5: Save the vectors to an image map database