| In recent years,with the advancement of artificial intelligence and deep learning technology,the field of unmanned driving technology has received widespread attention.However,applying unmanned driving technology to complex real-world traffic scenarios remains a challenge,and environmental perception systems play a crucial role in overcoming this challenge.The system is responsible for detecting and recognizing various elements on the road,such as lane lines,pedestrians,vehicles,traffic signs,and traffic lights,and determining the position of obstacles in the environment to ensure the safe operation of unmanned vehicles.For low-speed unmanned vehicles,traditional environmental perception algorithms cannot meet the needs of semantic extraction in open environments.Perception technology based on deep learning has completely changed the unmanned driving field,exceeding the limitations of traditional environmental perception algorithms.However,low-power embedded devices have become the main computing unit of unmanned driving systems,which cannot effectively support the real-time operation of large deep learning models.Therefore,to address these issues,this article constructs a visual environment perception system for low-speed unmanned vehicles,proposes a lightweight obstacle object detection,single-image depth estimation,and lane line detection multi-task network,improves the environmental perception ability of unmanned vehicles,and is of great significance to unmanned vehicle environmental information cognition and decision-making planning.The main research achievements of this article are described as follows:(1)In response to the problem that the computation amount of the object detection algorithm based on neural network leads to the slow feature extraction speed of single frame image and cannot run in real time in embedded system,this paper proposes a lightweight backbone feature extraction network GosCSPNet.On this basis,an end-to-end lightweight object detection model based on multi-scale feature fusion is proposed.To improve the accuracy of the model,this paper introduces the αIoU loss function to enhance the training effect of the model.The object detection model achieved 40.4mAP on the COCO2017 dataset with a computing consumption of only 2.7GFlops.In order to analyze the real-time performance of the model on actual devices,it was deployed and verified on the embedded device Jetson AGX Xavier.The experimental results show that compared with existing object detection algorithms,the proposed object detection model in this paper achieved the highest running speed without affecting the detection accuracy,reaching 54.1FPS on the device.(2)In response to the problem that existing object depth estimation models are complex and inaccurate,this paper proposes a deep learningbased multi-task object distance measurement fusion network,and uses sparse depth samples for auxiliary training.The experiment was verified on the KITTI monocular depth estimation dataset.The experimental results show that the δ1,δ2,and δ3 accuracy of the proposed monocular depth estimation method in this paper increased by 38.2%,4.1%,and 1.6%,respectively,compared with the baseline method.(3)In order to verify the overall performance of the proposed method,this paper uses the unmanned vehicle as the mobile platform,equipped with a monocular camera and embedded development board to carry out experimental tests in the actual campus and industrial park environment.In addition,in view of the low detection accuracy and robustness of the existing lane detection model for curves,this paper designed a strategy to build a training database of curve detection algorithm,so as to improve the accuracy and robustness of the lane detection model on curves.Experimental results show that the environmental sensing system achieves 91.4%accuracy and 10.1FPS on the curve data set,which proves the feasibility and robustness of the proposed algorithm in the low-speed unmanned vehicle operation and meets the task requirements in the actual scenario. |
PDF Full Text Request