Research On Intelligent Vehicle Environment Perception Method Integrating Binocular Vision

With the rapid development of artificial intelligence technology and the automotive industry,intelligent vehicles with advanced driver assistance system(ADAS)and autonomous driving technology as the core have gradually become the focus of future research and the new direction of the development of the automotive industry.Driving environment situational awareness is one of the core technologies of the intelligent vehicle automation system,and is also the basis of vehicle global or local path planning and horizontal and vertical control.Therefore,the accuracy of the perception system is improved by optimizing the core algorithm in the environment awareness task in intelligent driving scenarios Robustness is of great significance for vehicle decision control and driving safety.In the forward environment perception scheme of intelligent vehicles,the system configuration scheme based on the full vision sensor is more suitable than the lidar scheme or the multi-sensor fusion scheme because of its low cost,low equipment selection difficulty,and rich perception information.As a mass production program.However,the two-dimensional image based on the traditional monocular visual perception algorithm itself has lost the spatial physical scale,which makes the algorithm more difficult to achieve environmental depth measurement and obstacle movement situation awareness.Faced with the high dynamics and complexity of various elements of the driving environment Performance,existing algorithms generally suffer from poor robustness.In order to solve the inherent defects of monocular vision in depth perception tasks,binocular vision technology is used to effectively improve the accuracy and robustness of distance detection of environmental elements through the effective combination of visual measurement methods and parallax principles.The image data acquired by the camera is used to realize other perception tasks.Based on the binocular camera,the proposed fusion algorithm for binocular vision includes three core tasks: environmental depth estimation,target detection,and structured road lane marking detection.The main research content includes the following aspects:1)The binocular stereo matching algorithm in the environment depth estimation task optimizes the classic SGM(semi-global matching)algorithm,and the original algorithm is applied to the weak texture area,occlusion area,and strong exposure environment matching in the driving scene.For the problem of poor performance,a matching cost calculation method combining pixel gradient and AD-Census transform is proposed.The weighted least squares filtering algorithm is introduced in the disparity refinement stage to improve the adaptability of the original algorithm to complex scenes.Aiming at the problem that the distance measurement accuracy of binocular sensors decreases with distance,a fuzzy distance perception method that imitates human drivers is proposed based on the parallax depth map obtained by stereo matching,and the target bounding box information output by the target detection task is fused using the 3 s algorithm Single target depth correction.2)The target detection task uses deep convolutional neural network technology to propose a YOLO-Sign Net target based on the task decomposition mechanism for the problem of difficulty in optimizing classification loss and poor detection accuracy caused by the imbalance of multiple types of target data during the training phase of complex neural networks.Detection algorithm.This algorithm merges traffic flow control signals with complex semantic types into a unified parent class,uses the YOLO-V3 algorithm to detect the position of the parent class target,and then inputs the parent target interest area to the separately designed Sign Net traffic signal classification network to complete Recognize the semantics of traffic signal lights and traffic sign subclasses.Through task decomposition,individual control of the training process of complex categories of targets and targeted adjustment and structural optimization of sub-task network hyperparameters are achieved.3)The lane marking detection task is based on the Lane Net algorithm framework.The detection task is split into two parallel subtasks: semantic segmentation and lane line pixel instance feature clustering,and the lane line detection is redesigned based on the VGG16-FCN semantic segmentation model.The network optimizes and upgrades the original basic network ENet,thereby improving the detection accuracy of the model in complex road scenarios.In the instance feature clustering stage,the original Meanshift algorithm is replaced by the DBSCAN algorithm.At the same time,the accuracy and speed of clustering are improved by introducing skeleton extraction ideas in the clustering process,which makes the improved Lane Net obtain a more excellent than the original algorithm.performance.Finally,in a variety of typical autonomous driving scenarios,the task algorithm proposed in the environment perception method integrating binocular vision was experimentally verified,and the environment awareness software was designed to uniformly schedule different algorithm threads to achieve information interaction.