Research On Hospitability Map Building And Ground Target Tracking

Ground target tracking is one of the most critical issues in UAVs study. There are many factors which influence the accuracy of tracking such as the complexity of the ground situation, diversity of the ground targets, precision of the UAVs airborne sensors, etc. Since the movements of ground targets constrained by the actual ground situation on a large scale, the tactical UAVs can obviously use such information to achieve better detection and tracking performance from multiple perspectives. This dissertation presented a simultaneous hospitability map building and target tracking method, the UAVs build ground targets' hospitability maps through the integration of real-time access to their own reconnaissance of environmental information, and targets estimation and tracking are carried through Bayesian filtering methods.First, with an exhaustive analysis of the advantages and disadvantages of existing target tracking approaches, a generic method for target track maintenance, the BF-HMap approach, was proposed based on the Bayesian filtering method and the hospitability map. Due to the flexible scheme of Bayesian inference inherent in this algorithm, BF-HMap method is capable of exploiting non-analytic prior environmental knowledge as well as handling intermittent and regional measurements caused by the coverage and motion constraints on UAVs. Comprehensive performance comparison shows that BF-HMap algorithm estimates and tracks targets far better than existing methods under intermittent and regional observation.The second major part considers the automatic road extraction from UAV remote sensing images. Roads in the environment are regarded as the most important components of hospitability maps. In order to build hospitability map automatically, main roads need to be extracted from the UAV remote sensing images automatically. Proposed method is based on two previously developed approaches: the first one is the differential geometric approach to extract linear features; second one is road contour extraction based on Active Contour Models (also known as Snake). A model-based strategy is used to recognize main roads in images, whose extraction processes as follows: First, one-directional bottom-up methods were used to generate the hypotheses of roads in images. Then, the hypotheses were verified and classified with its return to the original date. Three road models were used to extract salient roads, non-salient roads, and road crossings, respectively.Further research work on the construction of hospitability map is given in the third part of this dissertation in three main aspects. First is to propose a D-S theory based hospitability map construction method for the fusion of information from multiple sensors. Second is to use Positioning Orientation System (POS) to realize direct geo-referencing. Third is to present a SIFT based automatic stitching of aerial image sequence method for the generation of global hospitability map. It consisted of two core steps: image registration and mosaic. As two consecutive frames in aerial image sequences often exhibit significant differences, the present commonly used feature-based methods are not very suitable for their registration. Concerned with this issue, an iterative registration method using the SIFT features was put forward to achieve accurate and robust aerial image registration. Experiments on simulated aerial sequences prove the effectiveness of the proposed method.Ground target tracking by UAVs was studied in this dissertation. A simultaneous hospitability map building and target tracking method was proposed to improve the UAVs target tracking performance in unknown environment. The proposed method fully considered the following issues: non-linear Bayesian filter design, the concept and construction of hospitability map, and remote sensing image interpretation and position, all of which constituted a set of relatively complete UAVs ground target tracking algorithm.