Research On Technologies Of Vision-based Structured Road And Obstacle Detection

Environmental perception is an important part in intelligent driving assistant systems,and it is also the core function of intelligent vehicles to perceive surrounding environment situation and its driving status.In the process of improving vehicle driving safety,most of environmental perception functions are based on vision because visual sensors are cheap,reliable and informative.The environmental perception ability of assistant driving system based on machine vision determines the ability of intelligent vehicle to avoid obstacles and to plan driving route in local areas.Its real-time performance and detection accuracy are the most important issues in current research.Aiming at the core problem of feature extraction and target detection in driving environmental perception,this thesis studies both traditional methods and deep learning.The main researches are as follows:In the part of lane line detection based on traditional methods,aiming at the problem that current traditional algorithm is sensitive to noise and environmental factors,and lane lines with large curvature can not be effectively fitted,this thesis proposes a real-time lane detection method based on hyperbolic model.In line extraction of near field,a new form of inverse perspective mapping is applied.After line segments detection from source image,these line segment coordinates are projected to the bird-eye plane by using the sub-pixel space mapping matrix between the perspective image and the top-view image,which effectively reduces interpolation error and time caused by generating bird-eye image.Secondly,this thesis uses DBSCAN to cluster lane edge segments and designs a fusion method to extract the best combination of clusters under specific constraints.In the hyperbolic curve fitting,this thesis proposes a weighted hyperbolic model to fit far-field lane lines.It can fit the curvature of lane lines and get more accurate results on lanes with large curvature.Because traditional methods focus on the features of lane lines,it is difficult or even impossible to segment lanes in the condition of lane lines incomplete or absence.Because some open source datasets mark lanes according to prior knowledge even if lanes are occluded or not existed,this thesis proposes an end-to-end convolution neural network to solve this problem.This thesis mainly uses the idea of semantic segmentation,and eliminates complex post processings.IBN normalization module and attention mechanism are used to optimize FCNVGG16 semantics segmentation network.A multi-task structure is used to obtain semantics segmentation feature maps and the instance segmentation feature map respectively.Finally,a simple cubic polynomial is used to fit lanes.The effectiveness of the proposed method is verified by comparison.Aiming at the problem of false alarms caused by departure from road surface in lane detection and achieving various objects detection ability in environmental perception,a lightweight and real-time road environment semantics segmentation network is proposed in this thesis.Although many methods based on convolution neural network have achieved good results on various data sets,the complexity and time overhead of the network make it impossible to be applied on vehicle platform.In this thesis,a new multi-scale parallel module based on Inception structure is proposed,which uses multi-scale convolution parallel connection on a single layer to refine egmentation accuracy under low overhead.The experimental results show that the proposed method can achieve real-time segmentation speed and guarantee the segmentation accuracy.Finally,traditional vehicle detection method based on vehicle platform is limited by sensors’ position and field of vision,so it can not detect vehicles effectively in the case of congested roads and vehicle occlusion.In order to solve this problem,an improved FT saliency vehicle detection method based on UAV platform is proposed.This method can still achieve real-time performance in complex environment.Secondly,this thesis combines Boolean map and OTSU to segment the region of interest(ROI)of vehicles on saliency images.This method can be used to solve time-consuming problem of Meanshift-based segmentation method.Finally,a series of vehicle visual feature recognition methods based on geometric,symmetric and horizontal edge waves are proposed to accurately identify vehicles and eliminate interference of roadside objects.The algorithm has been transplanted to UAV platform.The experimental results show that the method can effectively and reliably detect multiple targets in complex urban road environment.