Image registration and stitching algorithm of rice low-altitude remote sensing based on Harris corner self-adaptive detection.

Automation of images registration and stitching is one of the most important key technologies to the wide use of the low-altitude remote sensing by Micro-UAVs (unmanned aerial vehicles) in rice growing. In order to overcome the limitations, i.e. the thresholds need to be artificially determined for the traditional Harris corner detection algorithm, this paper proposed a self-adaptive algorithm for Harris corner detection, which was used in image registration and stitching of the rice low-altitude remote sensing. The algorithm was improved based on the traditional Harris corner detection algorithm by using a self-adaptive threshold determination method, which calculated from the standard deviation of image pixel gray-scale value. And then the characteristics of image were described by corners, and the images were registered by using the Euclidean distance among descriptors. In order to verify the effectiveness of the algorithm and optimize the relevant parameters, a verification test was conducted based on low-altitude remote sensing images, which were gained by a multispectral camera mounted on a multi-rotor unmanned helicopter during rice tillering stage. Four indices, the repetition rate (a measure of the stability of corner detection), the recognition rate (a measure of corner recognizable description operator), the registration rate (a measure of the accuracy of image registration and stitching) and running time of algorithm (a measure of computing speed of the algorithm), were proposed to evaluate the results of registration and stitching. Sixty images were randomly divided into 3 groups for verification test. Test results showed that the average registration rate reached 98.95%, and also the average repetition rate reached 96%, which indicated that the proposed algorithm had high accuracy. The repetition rate and the difference in image registration rates among the groups were not significant (at 0.05 significance level), which indicated that the proposed algorithm was stable and reliable. And the recognition rate among the groups was significant, and it indicated that the proposed algorithm had higher distinguishability to the remote sensing images, which was conducive to the precision of the automation of images registration and stitching. Threshold value of the proposed algorithm, which is the standard deviation of the image pixel gray values after standardization, here is set to 1 and 2 for optimization test. Test results showed that the registration rate was not significant, namely there was no significant difference (at 0.05 significance level) when the threshold value was equal to 1 or 2. However, comparing the average running time of the proposed algorithm, it showed that the running time when the threshold value was 1, is 2.5 times that when the threshold value was 2. Based on comprehensive consideration of the registration rate, the running time and the efficiency, the threshold value of 2 can be set as the optimum parameter of the proposed algorithm. [ABSTRACT FROM AUTHOR]