Research On Vision Perception Algorithms For Autonomous Vehicles On Highway Roads

Under the support of the emphasis projects from the Natural Science Foundation ofChina (NSFC):‘Research on key scientific problems in intelligent driving of highwayvehicles’(90820302) and ‘Research on theories and methods for navigation control ofmobile robots in unknown environments’(60234030), this thesis investigates the VisionPerception problems that Autonomous Land Vehicle (ALV) merges into traffic flow onHighway Roads.In the normal traffic environment on highway, to drive and merge into traffic flowsafely, the Autonomous Land Vehicle should satisfy the basic condition which is todetect reliably the road shape and the space relation between vehicle and road. Thevision perception system of the autonomous Hongqi HQ3is designed for some complexsituations such as unreliable lane marking lines. The main contributions and innovationsof this thesis are as follows:1. The consistency between the highway design specifications and the laws ofhuman visual perception has been proposed. The road details imply lots of informationfor visual perception of the autonomous driving system. Thus visual perception systemhas been proposed for complex road environment, for example, when the lane markinglines are faded, obscured, or missed.2. The mathematical meaning of “tangent point” focused on by drivers is to obtainthe curvature of the road. Using the coordinates of the Vanishing Point of the parallellines, this thesis estimates the curvature information of the road.3. The scene is classified into drivable terrain, vertical obstacle and sky. Four kindsof feature parameters are chosen, i.e., color, texture, position and perspective relation.The super-pixel method is adopted to segment the image into areas, greatly reducing thedata amount and improving the computing efficiency.4. The3D road curvature model is established based on road design and vehiclekinematics, and the profile curvature relating to the road undulation and the planecurvature relating to the road bending are analyzed, which can estimate the undulationand bending information of road in real time, and reduce the errors caused by the curvemodel.5. Using a serial of images and based the principle of “apparent movementphenomenon”, the problem of the vehicle trajectory estimation and the highway exitdetection has been solved effectively.The above results are core technologies of the lane perception system of theHongqi HQ3autonomous vehicle, they can reliably provide lane information, the spacerelationship between vehicle and road and the highway exits under the complex situations, such as the lane marking lines are faded, obscured, or missed. The processingcycle of this system is less than30milliseconds, which has satisfied requirement ofreal-time. This system can let an autonomous vehicle to merge into traffic flow at thespeed of110km/h on highway roads without Global Position System and inertialnavigation equipment. And it has provided reliable lane information guarantee for thefirst successful autonomous drive experiment on highway road with long-distance (286kilometers) under normal traffic conditions in China.