Research On Methods Of Moving Object Detection Based On Large Field Of View Camera In Vehicular Environment

As a great invention of human beings in the 20th century, intelligent vehicles have become the representative strategic goals of the countries in the high-tech field. Into the 21st century, governments, businesses and research institutions are vigorously promoting the practical progress of intelligent vehicles in military, civilian and scientific research fields. Moving object detection technology as one of the core technologies of intelligent vehicles is still a national study of key inputs. Especially for moving object detection in complex outdoor environment based on large field of view camera, there are still many difficult issues to be solved.The purpose of this study is to explore the theory and methods of moving object detection based on moving monocular camera of large field of view in vehicular environment. The main works of this dissertation include:Motion parameter estimation(1) It is very important to estimate motion parameters accurately when detecting moving objects. Based on flat road assumption, a motion parameter asynchronous estimation method which can applicable to fish-eye camera is proposed in this dissertation. The relations between motion parameters and displacement vectors of image points in different distances and positions are analysed and deduced. Based on the analyzing results, motion parameters are asynchronously estimated. It solves the problem in the previous algorithms that rotation parameters and translation parameters influence each other when estimating at the same time, and improves the estimation accuracy. Comparing to the traditional methods based on flat road assumption, the propesed method is more robust and accurate. The vehicle will generally bump in complex road environment and it will bring biggish error if using flat road assumption. Therefore, this dissertation extends the motion parameter asynchronous estimation method to the situation when the camera has three-dimensional motion to realize estimating three-dimensional rotation and translation parameters robustly and accurately Comparing to other ego-motion parameter estimation methods available to three-dimensional motion, this dissertation's method solved the interfering problem between parameters and reduced the parameters' estimation complicacy and obtained better estimation results.Moving object detection(2) Moving object detection method based on traditional focus of expansion (FOE) point of optical flows is not applicable to large field of view camera. This dissertation proposes an improved FOE method (GFOE) to solve this problem. Based on GFOE theory, in practical application, the GFOE method based on feature point clustering (CGFOE) is ulteriorly proposed which can effectively reduce the effect of feature points' incorrectly matching and ego rotation parameters' precision error. This dissertation also analyzes the principle applying situations of the GFOE method and gives the available range and limitis of this method. Experimental results show that the method has good detection effect for most moving objects, especially suitable for detecting the objects whose trajectories are crossing with the vehicle's trajectory.(3) Moving object detection method based on epipolar constraint is unavailable for detecting objects moving along the camera optical center's moving direction. So this dissertation proposes a moving object detection method based on reconstruction consistency to solve this problem. Firstly, a motion discriminant is derived by analyzing the geometry constraint of space point's imaging in adjacent frames. Then, the moving point clusters are detected by this discriminant and these clusters are used to get the finally detection results. The method gets good results. The dissertation also proposes a directly method to estimate time to collision (TTC) based on displacement vectors of feature points. The method doesn't need to recovery objects' depth information and it is available to estimate moving objects' collision time in the scene when the vehicle is reversing straightly.(4) Moving object detection method based on optical flows of feature points is easy to be effected by the points' error matching and is not so good for detecting objects those normally without available optical flows such as objects moving too fast or too small objects in distance, etc. As to such questions, the dissertation proposes a moving object detection method basd on multi-plane compensation. Comparing to the global motion compensation methods normally used, the multi-plane compensation method proposed in this dissertation can solve the compensation problem for the strong parallax backgrounds. Moreover, it is not necessary to use feature matching algorithm to solve model parameters, so it can forbid the error from matching. In order to increase the robustness of moving object detection method, the dissertation obtains the final moving object detection results by multi-frame fusion based on Bayesian theory. The experimental results show availability of the method.In conclusion, the proposed methods in this paper are effective to solve the problems in moving object detection based on large field of view camera in complex vehicular environment, and improve the precision of motion parameters estimation and the accuracy and robustness of moving object detection.