Research On Driving Environment Perception And Recognition Based On Computer Visio

Automatic driving technology is one of the most popular research directions in the field of deep learning.Among them,environmental perception during driving is one of the most critical tasks in automatic driving technology,including multiple perception tasks such as traffic target detection,drivable area segmentation,and lane detection.However,current perception algorithms usually only focus on one of these tasks,ignoring the correlation between different tasks and wasting the limited hardware resources of automatic driving vehicles,making it difficult to meet the needs of automatic driving vehicles to perceive multiple environmental factors simultaneously.To solve these problems,this paper proposes an end-to-end multi-task detection and segmentation network Md SNet(Multi-task detection and Segmentation network)model,which can achieve unified detection of traffic targets,drivable areas,and lane lines.The main research work of this paper is as follows:1.Multi-task model construction research.In terms of the backbone network,this paper draws on the CSPDark Net backbone network and combines the improved Res2 Net network to construct a multi-scale feature backbone network.In terms of feature fusion,this paper introduces the ULSAM attention mechanism to better integrate shallow detail features and high-level semantic information.In the target detection task processing stage,this paper uses the PAN network with up-aggregation structure and YOLO detection head to locate and recognize targets.In the drivable area segmentation and lane detection task stage,this paper draws on the idea of the ENet network to construct a fast segmentation network and adds the DLFM module to improve the model’s ability to extract drivable area and lane line boundary information.2.Multi-task model training method research.Firstly,the BDD100 K and Camvid datasets are preprocessed and data augmentation is performed.Secondly,multiple loss functions are designed to be automatically weighted and summed to optimize the performance of the entire network.At the same time,according to the characteristics of different target scales,specific Anchor mechanisms are designed to reduce regression difficulty and improve algorithm detection accuracy.Finally,specific model training strategies are designed to compress model training time.3.Multi-task model validation and evaluation.This article conducts extensive comparative and ablation experiments on the BDD100 K and Camvid datasets.The experimental results demonstrate that Md SNet achieves excellent detection performance on both datasets,with m APs of 47.7% and 58.2% for traffic object detection,m Io Us of 92.6% and 94.5% for drivable area segmentation,and Io Us of 62.6%and 75.1% for lane detection,respectively.Moreover,the proposed model has a small size of only 7.02 M,which can provide technical support for later deployment on embedded mobile terminals.