| With the rapid development of national science and technology and the improvement of people’s living standards,the demand for unmanned technology in the fields of social services,economic development,national defense construction and other fields is becoming more and more urgent.Unmanned driving visual perception is an important part of unmanned driving technology,so the research of unmanned driving visual perception is of great significance.It is mainly divided into two aspects,object detection and drivable area segmentation.Object detection includes vehicle detection,pedestrian detection,traffic sign detection,traffic light detection,etc.The drivable area segmentation includes direct and indirect drivable area segmentation,which is affected by vehicles,pedestrians,and other obstacles,and restricted by lane lines,traffic lights and traffic signs.In order to study the unified model and algorithm of object detection and drivable area segmentation,research on object detection algorithms based on two deep learning are presented in this thesis,including:In terms of object detection,two object detection algorithms based on deep learning are used to detect ten types of objects in the BDD100 K dataset.Firstly,the dataset is statistically analyzed and then cleaned and enhanced according to the problems found.Secondly,the object detection algorithm is optimized to improve the accuracy of the ten types of object to be detected.In terms of segmentation of drivable area,the object detection method is used to segment the drivable area instead of the commonly used semantic segmentation method.Firstly,the BDD 100 K pixel-level labels are transformed into the label of object detection,in other words,a calculus-like idea is adopted to cover a region of irregular polygons with multiple rectangular boxes,thereby obtaining several rectangular box labels.Secondly,the transformed labels are trained to obtain the model.Thirdly,the test set is tested to get a series of rectangular boxes.Finally,these rectangular boxes are synthesized into regions to achieve segmentation.In terms of lane line detection,the object detection algorithm based on deep learning is also used to detect lane lines.Firstly,it is necessary to convert the original data label of Bezier curve into the object detection frame label.The conversion idea is to draw the Bezier curve,take several points on the curve,and the coordinate data of two adjacent points is used as the coordinate data of the diagonal vertices of the lane line rectangle.Secondly,the converted annotations are trained to obtain a model through which the test set is detected to obtain rectangular frame that is the approximate position of the lane line.Finally,the frame is processed by related methods to find the lane line’s specific location,so as to realize the lane line detection.In terms of the unified model,the three research contents: object detection,drivable area segmentation and lane line detection are trained based on two deep learning object detection algorithms,so that a unified model is obtained,and data processing and parameter tuning are optimized to improve the accuracy of detection and segmentation.The experimental results show that the unified model algorithm in this thesis can satisfy the basic object detection,drivable area segmentation and lane line detection.The relevant methods and parameter optimization in this thesis can improve detection accuracy. |
