Vision-based Unmanned Vehicle Environment Perception

With the development of computer hardware and software technology,the development speed and level of artificial intelligence technology have reached an unprecedented height,and driverless cars have gradually appeared in the vision of the public.Intelligence is the general direction of the future car driving technology,and the environmental awareness technology of unmanned vehicle plays a decisive role in the autonomous driving technology.The environmental awareness part of the autonomous driving system usually needs to acquire a large amount of information about the surrounding environmental information,specifically including the location,speed and possible behavior of obstacles,navigable area,traffic rules and so on.Unmanned vehicles usually obtain information by integrating data from various sensors such as lidar and camera,and extract key and efficient environmental information,which is an important prerequisite for the correct decision-making of the automatic system.vision-based autonomous vehicle perception technology has become the focus of academic research in the field of autonomous driving.Algorithms based on visual environment in this thesis,the study and actual application problem,set up an associated framework of super-resolution algorithm based on end-to-end and uncertainty of target detection algorithm,the noise of image information are extracted by edge enhancement network edge character and the uncertainty existing in the goal and return to quantitative assessment,combining with the gaussian distribution,for the system to provide more accurate environment information,and guiding the unmanned vehicle in obstacle avoidance,decision-making,safe driving and other operations,and built an online environment awareness platform and hardware platform for outdoor experiments.The main work of this thesis is as follows:1.Image enhancement model of GAN and EEN network.The model consists of two parts:GAN and the edge-enhancement Network(EEN).Intermediate super-resolution Image(ISR)through confrontation training.on the edge enhanced subnetwork,the replacement of Dense Block with residual in residue Dense Blocks(RRDB)reduced the calculation amount,improved the efficiency,and improved the stability of EEN network.The edge information in ISR images was extracted by EEN network to eliminate the noise,and then the ISR images generated against the GAN network and the edge features extracted by EEN network were fused to obtain the super-resolution(SR)image,and the validity of the model was verified through experiments.2.An uncertainty detection model based on Gaussian distribution.This section presents a prediction methods of positioning uncertainty using gaussian distribution,through the output coordinate modeling of YOLOv3 target detection algorithm,and the gaussian parameters redesigned the return loss function,the model has the better robustness for noise data,increasing the detection speed and positioning accuracy.it is verified that the proposed scheme can quantitatively evaluate the reliability of target positioning by experiment.It significantly reduces False-Positive(FP)and increases True-Positive(TP),thus improving the accuracy.3.In this section,an end-to-end joint target detection model based on GAN-EEN and Gaussian distributions is proposed.Firstly,the super-resolution image is generated based on GAN-EEN network,and then image classification and target location are carried out based on gaussian mixture detection model.Two more efficient losses are proposed for generators:the perceived loss and the content loss(L1).The perception loss is calculated before the activation layer,and advanced features are extracted to effectively improve the weak supervision performance caused by sparse activation.This thesis designed and built the unmanned vehicle environment entity platform based on visual perception,composed of intelligent cars and Hi Lens Kit,at the same time,design a simple movement control strategy and lane line recognition algorithm,according to the environmental information is detected,the unmanned vehicles along the driving path motion controlled by decision-making,but also can complete other extensibility tasks according to the results of the target detection.