The Research On Key Technologies Of Visual Perception Of Driverless Road Environment

The development and popularization of automatic driving technology will completely change the way people travel,rebuild the traditional automobile industry,and can effectively reduce the occurrence of road traffic accidents,ease traffic congestion.It will improve the safety and intelligence of traffic operations,and has become a research hotspot in the global industry and academia.An automatic driving system usually includes multiple components such as environmental perception,decision making and execution,among which accurate,real-time and stable road environment perception technology is the most important and basic part of an automatic driving system.Various traffic elements including vehicles,pedestrians,drivable areas and traffic signs are located and identified on the road by the road environment perception technology to provide guidance information for subsequent driving decision-making.Therefore,the accuracy and efficiency of road environment perception directly affect the safety and reliability of automatic driving.This paper focuses on the two key tasks of traffic sign detection and recognition and road environment understanding in automatic driving road environment perception technology.The specific research contents and innovations are as follows:1)Aiming at the problems of low detection accuracy of small targets and uneven precision of different categories caused by long tail distribution in the existing road traffic sign detection tasks,a new traffic sign detection and recognition method is proposed from the two aspects of metric model and loss function.First,in order to improve the detection precision of small traffic signs,a scale-insensitive metric for small target detection is designed.Compared with the traditional metric,the proposed metric is more conducive to the generation of positive samples during the training phase and can be applied to datasets containing targets of different scales;Further,a equalized focal loss function is introduced to improve the detection accuracy of tail-category traffic signs by dynamically improving the balance of positive and negative samples of tail-category in training and increasing their proportion in the overall loss.The proposed traffic sign detection and recognition algorithm achieves 55.4% m AP on the tt100 k traffic sign detection dataset,and the m AP is improved by 5.5% compared with the baseline model,in which the precision of small target and tail-class traffic signs is improved more obviously,which proves the effectiveness of the proposed method.2)Aiming at the difficulty of balancing speed and accuracy in existing road environment segmentation models,a real-time road environment understanding model based on multi-feature selective fusion is proposed.First,considering that the feature fusion operation in the existing real-time segmentation methods ignores the difference of information between different features and the dynamics of requirements,a multi-feature selective fusion module is proposed,which can adaptively select and fuse the required information contained in different features;further,in order to jointly extract multi-scale and global context information in images to improve segmentation consistency,a context acquisition module is constructed,which can efficiently extract context information and only bring a small amount of extra computation consumption and memory usage,so as to ensure the inference efficiency of the model.Experiments on the City Scapes dataset show that the proposed method achieves high-speed and high-precision processing results with an accuracy of 77.1% m Io U and 37 FPS for input images with a resolution of 1024*2048,which meets the real-time requirements of automatic driving for perception tasks.3)The system development and key technology verification are carried out on the embedded intelligent computing hardware.Based on the embedded intelligent computing device NVIDIA Jeston Ag X Xavier,an autonomous driving road environment perception demonstration system is developed,the system used the Py Torch deep learning framework to deploy the above traffic sign detection and recognition model and road environment understanding algorithm,and Py QT is used to develop a visual graphical interactive interface to realize the deployment of the algorithm and the visual display of the perception results.The actual driving video in different scene is collected and tested by the system,the test results show that the proposed perceptual algorithm and system can effectively and stably perceive the road environment under the computing power of embedded hardware devices.