Research On Detection Technology Of Unstructured Road Drivable Area Based On Vision

With the rapid development of artificial intelligence and computer vision technology,autonomous driving technology has become a hot area of concern.Vision-based road drivable area detection technology plays a vital role in intelligent driving environment perception tasks,which can help autonomous vehicles achieve accurate path planning and driving decisions,and improve driving safety and stability.At the same time,in view of the characteristics of poor unstructured road conditions,unstable road surface conditions,and imperfect traffic signs and markings,new and higher requirements are put forward for the performance of the drivable area detection model.Therefore,vision-based drivable area detection in unstructured road environments is of great significance.This thesis takes the vision-based intelligent vehicle as the research object,and the unstructured road drivable area detection as the research goal,that is,the road area in front of the vehicle is segmented,and multiple targets in the front area are detected and tracked at the same time.Firstly,by using the deep neural network to extract image features,combined with object detection,semantic segmentation,multi-task learning technology,etc.,the road multi-object detection and unstructured road segmentation models are optimized and improved to improve the performance of the detection and segmentation models.Secondly,based on the idea of multi-task,a multi-task unstructured road segmentation and multi-target detection model is designed by sharing and merging the road multi-target detection model and the unstructured road drivable area segmentation model.Real vehicle experiments are carried out on the Kit driverless platform.The main research content of the thesis is as follows.(1)Road multi-target detection and tracking.For the multi-target detection task in unstructured roads,by comparing various target detection algorithms based on deep learning,it is proposed to choose the YOLOv5 network with high precision,fast detection speed and relatively mature as the basic network of multi-target detection,and The optimization method of lightweight backbone network is applied to improve the real-time and inference efficiency of object detection tasks on resource-constrained devices,and improve the applicability.On this basis,the ordinary feature pyramid structure is replaced by a two-way feature pyramid structure,and an additional weight is added to each feature fusion input to enhance the model’s ability to extract objects such as vehicles and pedestrians in unstructured roads.In addition,by adding an attention mechanism in the detection network to emphasize regions of interest while ignoring irrelevant backgrounds,the maximum detection performance improvement can be achieved with a small computational cost.At the same time,based on the position and state information of the target detected by YOLOv5 in this frame,the Kalman filter algorithm is used to estimate the position and state information of the target in the next frame,and the prediction result obtained by the Kalman filter algorithm is compared with the image of the next frame.Hungarian matching,enabling visual tracking of specific targets.The test results show that the improved YOLOv5 target detection algorithm can effectively improve the accuracy and real-time performance of road multi-target detection,and based on the target detection results,combined with Kalman’s target tracking algorithm,it can achieve continuous tracking of specific targets in video sequences.(2)Unstructured road segmentation.For the road segmentation task in unstructured roads,by studying a variety of semantic segmentation algorithms based on deep learning,it is proposed to choose the Deeplabv3+ network with higher precision,fewer model parameters,and more advanced network as the basic network for road segmentation.On this basis,two different backbone feature networks are provided to adapt to the detection requirements in different unstructured road scenarios,and at the same time,a new weighting parameter is introduced based on the cross-entropy loss function to achieve difficult and easy samples when training the model.and the balance of positive and negative samples to improve the training performance of the model.The test results show that the improved Deeplabv3+ algorithm can more accurately realize the precise segmentation of unstructured road areas.(3)Multi-task unstructured road drivable area detection.This thesis intends to design a vision-based multi-task unstructured road drivable area detection scheme,integrate unstructured road segmentation,road multi-target detection and tracking modules,share a feature extraction network to complete the forward propagation,and realize the perception of the front At the same time,it can obtain the target information of vehicles and pedestrians ahead.The experimental results show that the multi-task unstructured road drivable area detection model can collect the input pictures and video information of the camera in real time on the Apollo D Kit driverless platform of Nanjing Forestry University,and complete multi-target detection and unstructured road Segmentation tasks can accurately segment road areas and detect vehicles and pedestrian targets,and meet certain real-time requirements.The experimental results of the above work show that the road multi-target detection and tracking scheme and the unstructured road segmentation scheme in this thesis can meet the requirements of unstructured road drivable area detection,and the detection performance meets certain accuracy and real-time requirements,and meets the requirements of automatic Driving environment perception needs.