Cooperative object corner detection algorithm for visual landing of unmanned helicopter

It is difficult to detect the cooperative object corner of the real-time image shot by the visual navigation system of a carrier-based unmanned helicopter because of the serious distortion of scale and angle. Therefore we propose a partitioned and bidirectional corner matching algorithm based on the scale invariant feature transform (SIFT). We design an asymmetric cooperative object which comprises red back, green H target and triangle. We use the color information of the cooperative object to segment and cut it from its background and then extract the SIFT features of a reference cutting image and a real-time cutting image respectively. We use the partitioned and bidirectional matching algorithm to improve the real time and accuracy of the SIFT feature matching. Firstly, we calculate the gravity centers of the H target and the triangle in the reference cutting image and the real-time cutting image and the edge points of the H target which are closest to the gravity center of the triangle. We use the three pairs of matching points to calculate the rough affine models of the reference cutting image and the real-time cutting image. Secondly, we use the models to transform the SIFT features of the reference cutting image, thus obtaining the mapping points of the real-time cutting image. We partition the mapping points by taking each mapping point as the central point of the circle and the 1/4 width of the real-time cutting image as the radius. We only match the SIFT features that are inside the region of mapping points. We use the above method to partition the reference cutting image. Then we use the bidirectional matching and the random sample consensus (RANSAC) algorithm to eliminate the wrong matching pairs. Thus we use the correct matching pairs to calculate the accurate affine model for the transformation between reference cutting image and real-time cutting image. Finally, the position of the cooperative object corner in the real-time cutting image is obtained by transforming the correct position of the cooperative object corner of the reference cutting image with the accurate affine model. The experimental results, given in Fig. 3 and Tables 1, 2 and 3, show preliminarily that our cooperative object corner detection algorithm is accurate, robust and real-time.