The world's first open source vehicle-road collaboration dataset is released, and data becomes the key to the development of autonomous driving

Those who pay attention to autonomous driving know that there are two major technical routes for autonomous driving, namely single-vehicle intelligent autonomous driving (AD) and vehicle-road collaborative autonomous driving (VICAD).

Single-vehicle intelligence has problems such as blind spots in driving and unstable perception at medium and long distances, which has limited the operational design domain (ODD) of autonomous vehicles and has encountered some obstacles in its implementation. Vehicle-road collaborative autonomous driving has begun to be seen as a trend for large-scale commercial implementation.

High-quality data is the key to vehicle-road collaborative autonomous driving technology. Faced with the lack of relevant data sets, it cannot meet the practical needs of all parties to achieve data-driven vehicle-road collaboration.

On February 24, the world's first real-scene based vehicle-road collaborative autonomous driving dataset DAIR-V2X was officially released and available for download and use by domestic users.

The dataset was jointly released by Tsinghua University's Institute of Intelligent Industries (AIR), Beijing High-Level Autonomous Driving Demonstration Zone, Beijing AutoNet Technology Development Co., Ltd., Baidu Apollo, and Beijing Zhiyuan Artificial Intelligence Research Institute.

The world's first open source vehicle-road collaboration dataset

The DAIR-V2X vehicle-road collaborative dataset is the first large-scale, multi-modal, multi-perspective dataset used for vehicle-road collaborative autonomous driving research. All data is collected from real scenes and includes 2D&3D annotations.

The dataset comes from 10 kilometers of real urban roads, 10 kilometers of highways, and 28 intersections in Beijing's high-level autonomous driving demonstration zone, and includes multiple types of sensors such as vehicle-side and road-side cameras and vehicle-side and road-side lidars.

Specifically, it includes 71,254 frames of image data and 71,254 frames of point cloud data, covering a variety of scenes such as sunny days, rainy days, foggy days, day and night, urban roads and highways.

Compared with datasets that only contain a single vehicle or a single road, this dataset provides multimodal data from the joint perspective of the vehicle and the road in the same time and space, and provides fusion annotation results from the joint perspective of different sensors, which can better serve the research and evaluation of vehicle-road collaborative algorithms.

In addition, the dataset effectively reduces the cost of dataset construction through innovations such as semi-automatic self-learning vehicle-road collaborative 3D fusion annotation methods.

The release of DAIR-V2X, the world's first autonomous driving vehicle-road collaborative dataset, is of great significance for promoting the research and development of high-level autonomous driving technology in my country.

At present, the dataset has been included in the Zhiyuan platform. In the future, it will rely on the Zhiyuan community and other Zhiyuan academic ecological networks to accelerate the opening, promotion and application of the dataset to all parties in industry, academia, research and application.

The large-scale commercialization of vehicle-road cooperative autonomous driving is promising

In February 2020, the National Development and Reform Commission, together with relevant ministries and commissions, issued the "Intelligent Vehicle Innovation and Development Strategy", which identified "promoting the planning and construction of intelligent road infrastructure" as an important national strategic task, and clarified the Chinese-style autonomous driving route of "single-vehicle intelligence + vehicle-road collaboration". Vehicle-road collaboration has become a research focus for all sectors.

Some people believe that single-vehicle intelligent autonomous driving still faces challenges and problems in terms of safety, ODD restrictions and economy in order to achieve large-scale commercialization. Under the current conditions of autonomous driving capabilities, it is impossible to find a balance between safety, ODD restrictions and economy, and it is necessary to fundamentally improve the capabilities of autonomous driving.

Vehicle-road collaborative autonomous driving can greatly expand the perception range of a single vehicle and improve its perception capabilities through information interaction, collaborative perception, and collaborative decision-making control, introduce new intelligent elements represented by high-dimensional data, and realize group intelligence.

It can fundamentally solve the technical bottlenecks encountered by single-vehicle intelligent autonomous driving, improve autonomous driving capabilities, thereby ensuring autonomous driving safety and expanding autonomous driving ODD.

Baidu founder, chairman and CEO Robin Li recently stated that using vehicle-road collaboration solutions for autonomous driving is the technical route that Baidu insists on and is optimistic about.

According to Baidu's deduction, the accident rate of autonomous driving can be reduced by 99% through vehicle-road collaboration; transferring some autonomous driving functions to the road side can greatly reduce costs, thereby accelerating the pace of commercialization of autonomous driving.