Scene Classification and Landmark Spotting in Multilayer Image Mosaic for UAVs in GPS-denied Situations

Abstract

In GPS-denied UAV navigation, it is important for the UAV system to autonomously adapt itself to the changing environment by analyzing sensor signals on the fly. The autonomous navigation system should analyze the instantaneous profile of the sensor data as well as manage the longer-term temporal evolution patterns of the environment. This requires the UAV system achieve the situational awareness at sensor data level and at contextual knowledge level in order to assist the autonomous decision making in the GPSdenied situation. The proposed system fulfills the requirements as follows. Firstly, the system builds the 'image hierarchy' incrementally on the fly by generating a multilayer image mosaic composed of 'short-duration mosaic' and 'long-duration mosaic' simultaneously. The 'long-duration mosaic' covers the overall area of the scene swept through by the UAV, while the 'short-duration mosaic' maintains and updates more detailed local area of the scene. Secondly, the system analyzes the individual image frame in terms of Hough line profiles, and classifies the scene into the 'man-made zone' versus the 'natural fields' categories. Each category is divided into two sub-categories. The Hough line profiles of the image frames show distinctive characteristics among the four sub-categories. The Hough line profile distribution is modeled by Gaussian mixtures, and the Bayesian classifier recognizes the scene categories of new image frames. Thirdly, the proposed system can select a specific landmark on demand such as a building or a crossroad from the input image and spot the corresponding landmark from the stored reference database such as satellite imagery. The proposed system achieves correct spotting of man-made objects between the UAV-acquired input image and the reference imagery database from satellites. The proposed system was evaluated with extensive field test data. The performances of the hierarchical multilevel mosaicing, the scene classification, and the landmark spotting show promising results.

A robust iterative super-resolution mosaicking algorithm using an adaptive and directional Huber-Markov regularization

“…Over the last decade, a number of algorithms related to mosaicking [5][6][7][8][9][10][11][12][13][14][15] and super-resolution reconstruction [16][17][18][19][20][21][22][23][24][25][26][27] have been proposed. Super-resolution reconstruction performance has been most influenced by the choice of ML estimation and regularization technique.…”