


Synchronized driving data of vehicle-road collaboration
Autonomous driving V2X-AD (Vehicle- to-everything-aided autonomous driving) has great potential in providing safer driving strategies. Researchers have conducted a lot of research on the communication and communication aspects of V2X-AD, but the effect of these infrastructure and communication resources in improving driving performance has not been fully explored. This highlights the need to study collaborative autonomous driving, that is, how to design efficient information sharing strategies for driving planning to improve the driving performance of each vehicle. This requires two key basic conditions: one is a platform that can provide a data environment for V2X-AD, and an end-to-end driving system with complete driving-related functions and information sharing mechanisms. In terms of the platform that provides the data environment, it can be achieved by utilizing the communication network between vehicles and the support of the infrastructure. In this way, vehicles can share real-time and environmental information needed for driving, thereby improving driving performance. On the other hand, end-to-end driving systems need to have complete driving functions and be able to share information. This means that the driving system should be able to obtain driving-related information from other vehicles and infrastructure and combine this information with its own driving planning to provide more efficient driving performance. While achieving these two basic conditions, security and privacy protection also need to be considered. Therefore, when designing the driving planning strategy of V2X-AD, we should pay attention to the efficiency of the information sharing strategy and thereby improve the driving performance of each vehicle. To sum up, vehicle-road collaborative assisted autonomous driving V2X-AD has huge potential
" For this reason, researchers from Shanghai Jiao Tong University and Shanghai Artificial Intelligence Laboratory published a new research article "Towards Collaborative Autonomous Driving: Simulation Platform and End-to-End System" proposes CoDriving: an end-to-end collaborative driving system that uses an information sharing strategy for driving planning to achieve efficient communication and collaboration. A simulation platform V2Xverse was built, which provides a complete training and testing environment for collaborative driving, including the generation of vehicle-road collaborative driving data sets, the deployment of full-stack collaborative driving systems, and closed-loop driving performance evaluation and driving tasks in customizable scenarios. Evaluation. "
At the same time, the simulation platform V2Xverse integrates the training and deployment test codes of multiple existing collaborative sensing methods, using a variety of test tasks to test comprehensive driving capabilities: 3D target detection, path planning, and loop closure. Autopilot. V2Xverse breaks through the limitations of existing collaborative sensing methods that can only "see" but not "control". It supports embedding existing collaborative sensing methods into a complete driving system and testing driving performance in a simulation environment. The researchers of this article believe that this will bring better functional extensions and a test benchmark that is more suitable for actual driving scenarios for vision-based vehicle-road collaboration research in autonomous driving.
- Paper link: https://arxiv.org/pdf/2404.09496
- Code link: https://github.com/CollaborativePerception /V2Xverse
Research background and significance
The research of this article focuses on collaborative autonomous driving based on V2X (Vehicle-to-everything) communication. Compared with single-vehicle autonomous driving, collaborative autonomous driving improves vehicle perception and driving performance through information exchange between the vehicle and the surrounding environment (such as roadside units, pedestrians equipped with smart devices, etc.), which will benefit people with limited vision. Safe driving in complex scenarios (Figure 1).
Figure 1. Dangerous "ghost probe" scene, the bicycle cannot sense the occluded object
Currently, V2X-based vehicle-road collaborative work mostly focuses on optimizing module-level perception capabilities. However, how to use cooperative sensing capabilities to improve the final driving performance in integrated systems is still underexplored.
In order to solve this problem, this article aims to expand the collaborative sensing capability into a collaborative driving system covering comprehensive driving capabilities, including key modules such as perception, prediction, planning and control. Achieving collaborative autonomous driving requires two key foundations: a platform that can provide a data environment for V2X-AD; and the second is an end-to-end driving system that integrates complete driving-related functions and information sharing mechanisms. From a platform perspective, this work builds V2Xverse, a comprehensive collaborative autonomous driving simulation platform that provides a complete process from the generation of vehicle-road collaborative driving data sets to the deployment of full-stack collaborative driving systems and closed-loop driving performance evaluation. From the perspective of the driving system, this article introduces CoDriving, a new end-to-end collaborative driving system that designs and embeds a V2X communication-based collaboration module in a complete autonomous driving framework to improve collaborative driving performance by sharing sensory information. . The core idea of CoDriving is a new information sharing strategy for driving planning, which uses spatially sparse but important visual feature information for driving as communication content to optimize communication efficiency while improving driving performance.
V2Xverse: Vehicle-road collaborative driving simulation platform
The key feature of the V2Xverse proposed in this article is the ability to achieve offline benchmark generation of driving-related subtasks and in different scenarios Online closed-loop evaluation of driving performance fully supports the development of collaborative autonomous driving systems. In order to create a V2X-AD scene, V2Xverse sets up multiple smart cars equipped with complete driving capabilities in the scene, and places roadside units on both sides of the road through certain strategies to provide supplementary vision for the smart cars. In order to support the development of collaborative autonomous driving methods, V2Xverse first provides (vehicle-vehicle) and (vehicle-roadside unit) communication modules, and provides complete driving signals and expert annotations for system training, and also provides closed-loop driving evaluation A variety of dangerous scenarios. The simulation platform framework is shown in Figure 2.
Figure 2. V2Xverse simulation platform framework
Compared with the existing Carla-based autonomous driving simulation platform, V2Xverse has three advantages. First of all, V2Xverse supports multi-vehicle driving simulation, while the mainstream carla-leaderboard and its derivative platforms only support single-vehicle driving simulation. Second, V2Xverse supports full driving function simulation, while the existing collaborative perception simulation platform only supports functions related to the perception module. Third, V2Xverse supports comprehensive V2X-AD scenarios, including diverse sensor devices, model integration and flexible scenario customization; see Table 1.
Table 1. Comparison between V2Xverse and existing Carla-based autonomous driving simulation platform
CoDriving: End-to-end self-driving model for efficient collaboration
CoDriving consists of two components (see Figure 3): 1) End-to-end single-vehicle autonomous driving network, which converts sensor inputs into driving control signals; 2) Driving-oriented collaboration, where collaborators share information critical to driving Perceptual features are used to achieve efficient communication, and the perceptual features of bicycle BEVs are enhanced through feature aggregation. The enhanced perceptual features will help the system produce more accurate perceptual identification results and planning prediction results.
Figure 3. The overall framework of CoDriving
End-to-end autonomous driving network
The end-to-end single-vehicle autonomous driving network is based on the Modal inputs are used to learn output waypoint predictions, and a control module converts the waypoints into driving control signals. To achieve this, CoDriving integrates the modular components required for driving into an end-to-end system, including 3D object detectors, waypoint predictors and controllers. CoDriving uses Bird's Eye View (BEV) representation because it provides a unified global coordinate system, avoids complex coordinate transformation, and better supports collaboration based on spatial information.
Driving-oriented collaboration strategy
V2X collaboration solves the inevitable problem of limited visibility of bicycles through information sharing. In this work, this paper proposes a new driving-oriented collaboration strategy to simultaneously optimize driving performance and communication efficiency. The scheme includes i) driving intention-based perception communication, where CoDriving exchanges spatially sparse but driving-critical BEV perception features through a driving request module; and ii) BEV feature enhancement, where CoDriving utilizes the received feature information to enhance the performance of each collaborative vehicle. BEV perception characteristics. The enhanced BEV features will help the system produce more accurate perception recognition results and planning prediction results.
Experimental results
Using the V2Xverse simulation platform, this article tests the performance of CoDriving on three tasks: closed-loop driving, 3D target detection, and waypoint prediction. In the key closed-loop driving test, compared with the previous single-vehicle end-to-end autonomous driving SOTA method, CoDriving's driving score significantly improved by 62.49%, and the pedestrian collision rate dropped by 53.50%. In the target detection and waypoint prediction tasks, CoDriving performs better than other collaborative methods, as shown in Table 2.
Table 2. CoDriving is better than SOTA's single driving method in the closed-loop driving task, and is better than other collaborative sensing methods in the modular perception and planning subtasks
This article At the same time, the collaborative performance of CoDriving under different communication bandwidths was verified. In the three tasks of closed-loop driving, 3D target detection, and waypoint prediction, CoDriving outperformed other collaboration methods under different communication bandwidth restrictions, as shown in Figure 4.
Figure 4. Collaboration performance of CoDriving under different communication bandwidths
Figure 5 shows a driving case of CoDriving in the V2Xverse simulation environment. In the scene in Figure 5, a pedestrian in the blind spot suddenly rushed out of the road. It can be seen that the autonomous driving bicycle had a limited field of vision and was unable to avoid the pedestrian in advance, causing a serious car accident. CoDriving uses the shared vision characteristics of roadside units to detect pedestrians in advance and avoid them safely.
Figure 5(1). Compared to bicycle self-driving with limited vision, CoDriving uses the information provided by the roadside unit to detect pedestrians in the blind spotFigure 5 (2). CoDriving successfully avoided pedestrians, but the self-driving bicycle did not avoid the situation in time, causing a collision
Summary
This work helps collaborative autonomous driving by building a simulation platform V2Xverse method development and proposes a new end-to-end self-driving system. Among them, V2Xverse is a V2X collaborative driving simulation platform that supports closed-loop driving testing. This platform provides a complete development channel for the development of collaborative autonomous driving systems with the goal of improving final driving performance. It is worth mentioning that V2Xverse also supports the deployment of a variety of existing single-vehicle autonomous driving systems, as well as the training and closed-loop driving testing of a variety of existing collaborative sensing methods. At the same time, this paper proposes a new end-to-end collaborative autonomous driving system CoDriving, which improves driving performance and optimizes communication efficiency by sharing key driving perception information. A comprehensive evaluation of the entire driving system shows that CoDriving is significantly better than the single-vehicle self-driving system in different communication bandwidths. The researchers of this article believe that the V2Xverse platform and CoDriving system provide potential solutions for more reliable autonomous driving.
The above is the detailed content of Open source! V2Xverse: handed over and released the first simulation platform and end-to-end model for V2X. For more information, please follow other related articles on the PHP Chinese website!

1 前言在发布DALL·E的15个月后,OpenAI在今年春天带了续作DALL·E 2,以其更加惊艳的效果和丰富的可玩性迅速占领了各大AI社区的头条。近年来,随着生成对抗网络(GAN)、变分自编码器(VAE)、扩散模型(Diffusion models)的出现,深度学习已向世人展现其强大的图像生成能力;加上GPT-3、BERT等NLP模型的成功,人类正逐步打破文本和图像的信息界限。在DALL·E 2中,只需输入简单的文本(prompt),它就可以生成多张1024*1024的高清图像。这些图像甚至

“Making large models smaller”这是很多语言模型研究人员的学术追求,针对大模型昂贵的环境和训练成本,陈丹琦在智源大会青源学术年会上做了题为“Making large models smaller”的特邀报告。报告中重点提及了基于记忆增强的TRIME算法和基于粗细粒度联合剪枝和逐层蒸馏的CofiPruning算法。前者能够在不改变模型结构的基础上兼顾语言模型困惑度和检索速度方面的优势;而后者可以在保证下游任务准确度的同时实现更快的处理速度,具有更小的模型结构。陈丹琦 普

Wav2vec 2.0 [1],HuBERT [2] 和 WavLM [3] 等语音预训练模型,通过在多达上万小时的无标注语音数据(如 Libri-light )上的自监督学习,显著提升了自动语音识别(Automatic Speech Recognition, ASR),语音合成(Text-to-speech, TTS)和语音转换(Voice Conversation,VC)等语音下游任务的性能。然而这些模型都没有公开的中文版本,不便于应用在中文语音研究场景。 WenetSpeech [4] 是

由于复杂的注意力机制和模型设计,大多数现有的视觉 Transformer(ViT)在现实的工业部署场景中不能像卷积神经网络(CNN)那样高效地执行。这就带来了一个问题:视觉神经网络能否像 CNN 一样快速推断并像 ViT 一样强大?近期一些工作试图设计 CNN-Transformer 混合架构来解决这个问题,但这些工作的整体性能远不能令人满意。基于此,来自字节跳动的研究者提出了一种能在现实工业场景中有效部署的下一代视觉 Transformer——Next-ViT。从延迟 / 准确性权衡的角度看,

3月27号,Stability AI的创始人兼首席执行官Emad Mostaque在一条推文中宣布,Stable Diffusion XL 现已可用于公开测试。以下是一些事项:“XL”不是这个新的AI模型的官方名称。一旦发布稳定性AI公司的官方公告,名称将会更改。与先前版本相比,图像质量有所提高与先前版本相比,图像生成速度大大加快。示例图像让我们看看新旧AI模型在结果上的差异。Prompt: Luxury sports car with aerodynamic curves, shot in a

人工智能就是一个「拼财力」的行业,如果没有高性能计算设备,别说开发基础模型,就连微调模型都做不到。但如果只靠拼硬件,单靠当前计算性能的发展速度,迟早有一天无法满足日益膨胀的需求,所以还需要配套的软件来协调统筹计算能力,这时候就需要用到「智能计算」技术。最近,来自之江实验室、中国工程院、国防科技大学、浙江大学等多达十二个国内外研究机构共同发表了一篇论文,首次对智能计算领域进行了全面的调研,涵盖了理论基础、智能与计算的技术融合、重要应用、挑战和未来前景。论文链接:https://spj.scien

译者 | 李睿审校 | 孙淑娟近年来, Transformer 机器学习模型已经成为深度学习和深度神经网络技术进步的主要亮点之一。它主要用于自然语言处理中的高级应用。谷歌正在使用它来增强其搜索引擎结果。OpenAI 使用 Transformer 创建了著名的 GPT-2和 GPT-3模型。自从2017年首次亮相以来,Transformer 架构不断发展并扩展到多种不同的变体,从语言任务扩展到其他领域。它们已被用于时间序列预测。它们是 DeepMind 的蛋白质结构预测模型 AlphaFold

说起2010年南非世界杯的最大网红,一定非「章鱼保罗」莫属!这只位于德国海洋生物中心的神奇章鱼,不仅成功预测了德国队全部七场比赛的结果,还顺利地选出了最终的总冠军西班牙队。不幸的是,保罗已经永远地离开了我们,但它的「遗产」却在人们预测足球比赛结果的尝试中持续存在。在艾伦图灵研究所(The Alan Turing Institute),随着2022年卡塔尔世界杯的持续进行,三位研究员Nick Barlow、Jack Roberts和Ryan Chan决定用一种AI算法预测今年的冠军归属。预测模型图


Hot AI Tools

Undresser.AI Undress
AI-powered app for creating realistic nude photos

AI Clothes Remover
Online AI tool for removing clothes from photos.

Undress AI Tool
Undress images for free

Clothoff.io
AI clothes remover

AI Hentai Generator
Generate AI Hentai for free.

Hot Article

Hot Tools

mPDF
mPDF is a PHP library that can generate PDF files from UTF-8 encoded HTML. The original author, Ian Back, wrote mPDF to output PDF files "on the fly" from his website and handle different languages. It is slower than original scripts like HTML2FPDF and produces larger files when using Unicode fonts, but supports CSS styles etc. and has a lot of enhancements. Supports almost all languages, including RTL (Arabic and Hebrew) and CJK (Chinese, Japanese and Korean). Supports nested block-level elements (such as P, DIV),

Safe Exam Browser
Safe Exam Browser is a secure browser environment for taking online exams securely. This software turns any computer into a secure workstation. It controls access to any utility and prevents students from using unauthorized resources.

MinGW - Minimalist GNU for Windows
This project is in the process of being migrated to osdn.net/projects/mingw, you can continue to follow us there. MinGW: A native Windows port of the GNU Compiler Collection (GCC), freely distributable import libraries and header files for building native Windows applications; includes extensions to the MSVC runtime to support C99 functionality. All MinGW software can run on 64-bit Windows platforms.

Notepad++7.3.1
Easy-to-use and free code editor

SublimeText3 Linux new version
SublimeText3 Linux latest version
