Research On Key Techniques For Intelligent Vehicles Vision/Radar Navigation System

Intelligent vehicle navigation system is an very important part of ITS (Intelligent Transportation System), and the research on key techniques for intelligent vehicle navigation system does great contribution to safe driving and even.In this paper, with the help of information processing method such as image analysis and information fusion, considerable research on some key techniques for intelligent vehicle navigation system, which include the methods for road detection on the monocular vision, preceding car detection on the monocular vision and interval detection by two sensors fusion, is done, and several effective and simple algorithms is proposed accordingly. What's more, an emulation software has been designed to test those algorithm. The most distinctive parts would be described in the following aspects:1. A two-step method based on image processing technique and optimization algorithm is proposed to detect lane in structural road environment. The method uses not only the local characters but the global characters of the input images to recognize lane more robustly, and divides every image to two portions in which the regression is processed respectively. With appending appropriate factors, the optimization algorithm can get more antinoise capability.2. An improved monocular vision method is studied for intelligent vehicle to detect the preceding car in structural road environment in the daytime. Through identifying the edges of the car, the object is detected; the false object is eliminated and the eligible object expressed as a 2-D model is acquired. Then the location of the object in the next frame is predicted by Kalman filter, and the object is detected near that location. Through the likelihood function we defined, the likelihood of the object is calculated. If the likelihood is too low, the object is detected over again, or else the 2-D model is updated.3. A monocular vision method is put forward for intelligent vehicle to detect the preceding vehicle at night, in which the information of the rear lights is used. Through the double threshold method, the distinct features from different images are combined and the rear light candidates are decided. Then a match function is defined, and the candidates are paired. After that, the location of the object in the next frame is predicted by Kalman filter, and a pair of rear lights are detected near that location. The result is validated by the symmetry of the rear lights. If the result is not satisfied.the object is detected over again.4. A two-level algorithm for vision/radar fusion is studied. In the decision level fusion, the basic belief assignment (BBA) of D-S evidence reasoning theory is designed, and a spatio- temporal fusion method is applied. In the feature level fusion, an adaptive weighted mean method is proposed.5. A simulation system for virtual testing is developed. Designed with the help of OpenGL graphical interface software, it is a vivid emulator for sensor fusion, in which vision information and synchronous radar information are simulated to testify the fusion algorithm proposed easily.6. The horizontal offset, yawing angle and interval are calculated, and according to the driving safety model, a decision method to judge the warning level for driver is discussed. And then a vision-based lane departure and rear-end collision warning system in intelligent vehicle is developed in this paper. According to the situation, distinct orders are given to the driver.Further more, the intersection of different algorithm is considered in this paper, such as the combination of road detection and preceding car detection, and the automatic switch between car detection in the daytime and at night.