Abstract: The fixed threshold selection of traditional ORB algorithm results in many false extractions
and mismatches, which cannot meet the requirements of accurate extraction and matching of different
image feature points. To solve this problem, an improved ORB feature point extraction and matching
method was proposed. Firstly, the local adaptive threshold was set up. Then, an adaptive threshold
selection criterion was designed by classifying the pixels, and thus the precise extraction of ORB
feature points was achieved. Finally, the PROSAC algorithm was used to complete the matching of
feature points based on the improved ORB feature points. The experimental results indicate that the
improved method has a high adaptability to variations in brightness, and both the calculation speed
and extraction accuracy are greatly improved. The total matching time is reduced, the number of
mismatches is less, and the accurate matching rate is increased, which indicates that this improved
method is characterized with accuracy and real-time performance. In addition, the RMSE error
obtained by tracking the feature points acquired at the matching stage is small, which demonstrates a
significant improvement in matching accuracy. Compared with other existing methods, this method
has better environmental adaptive capacity and application value.

Key words: feature point extraction, local adaptive threshold, repetition rate, point pairs matching;
tracking