| Machine vision has become the first choice of many solutions in the autonomous driving environment perception system.However,it still has problems such as poor adaptability to complex driving environment and high requirement of calculation force.Meanwhile,its intelligence is still not up to the general level of human vision.How to simulate human vision to understand the complex driving scenarios better has become one of the key issues to promote the technological change of ADAS(Advanced Driving Assistant System)and realize the L4-level fully visual perception on the autonomous driving technology.Aiming at meeting the requirements for the optimal speed,accuracy and robustness of machine vision inspection tasks in complex driving scenarios,this paper takes the driver’s visual attention,road objects and drivable area as the research objects to carry out the research on a multi-task detection with human visual imitation.The main work of this paper is as follows:(1)Due to the poor accuracy caused by the lack of attention screening mechanism for complex driving scenarios in the existing object detection algorithms of machine vision,it is proposed to utilize the driver’s high-level perception feature to detect multiple types of objects on the road.Based on attention selectivity mechanism unique to human vision,it builds a model for the visual saliency of driving scenes and analyzes at the algorithm level.It is described from two aspects including the driving task-driven and the driver’s subjective awareness-driven.In addition,the driving visual attention salience dataset is constructed.(2)Due to the limited computing resources of on-board hardware and the difficulty and poor performance of multiple visual detection tasks performed by a single model in the existing visual detection algorithms,it proposes an end-to-end multi-task detection algorithm for complex driving scenarios.To improve the accuracy and robustness of the visual detection task,the excellent monocular machine vision algorithms based on deep learning are studied,including the road object detection algorithm based on anchor free and the drivable area detection algorithm based on "encoder decoder" structure.Then the design of the overall network structure and loss function is completed by combining the hard parameter sharing and soft parameter sharing mechanism in multi-tasking learning and the human visual selectivity mechanism.The overall structure is composed of a main module and three branch modules as well.The main module adopts the improved lightweight feature extraction network Mobile Net V3 as a backbone to realize the underlying feature sharing,which can ensure the real-time performance of the model.The three branch modules consist of the driver attention supervision module,the road object detection module and the drivable area detection module.The loss function is calculated by weighted joint method.(3)Establish an experimental platform to conduct offline training and online testing for the overall model.In addition,the algorithm was verified online on a laptop computer.The experimental results show that the model integrates the driver’s attention,which contributes the most to the mean Average Precision(m AP)value of road object detection increased by 3.65%.From the perspective of a single task of the model,compared with FCOS model,the parameters of detection part are only 8.79% and the Average Precision(AP)value of car category is improved by 0.56%.The detection time is also reduced by 8.55 ms.On the other hand,the accuracy of segmentation part is better than both UNet model and Seg Net model.Besides,its MAE value is the lowest.The multi-task model can accurately and steadily detect drivable areas and the road objects with three categories including car,rider and person in a real-time unified way,no matter in the sunny day or in the night under the condition of strong light intervention and backlight,which can meet the real-time requirements of normal on-board computing platform. |
PDF Full Text Request