Horizon releases open-source Sparse4D algorithm, pushing one step closer to end-to-end autonomous driving

On January 22, Horizon will open source the Sparse4D series of pure visual autonomous driving algorithms to encourage more developers in the industry to participate in the exploration of cutting-edge technology directions such as end-to-end autonomous driving and sparse perception. Currently, the Sparse4D algorithm has been launched on the GitHub platform. Developers can follow the Horizon GitHub official account "Horizon Robotics" to obtain the source code.

Sparse4D is a series of algorithms towards long-term sparse 3D target detection, which belongs to the category of temporal multi-view fusion sensing technology. Facing the industry development trend of sparse perception, Sparse4D has built a pure sparse fusion perception framework to make the perception algorithm more efficient and precise, and to make the perception system simpler. Compared with the dense BEV algorithm, Sparse4D reduces the computational complexity, breaks the limitation of computing power on the perception range, and surpasses the dense BEV algorithm in terms of perception effect and reasoning speed. In both the nuScenes pure visual 3D detection and 3D tracking lists, Sparse4D ranked first, becoming SOTA, ahead of the latest methods including SOLOFusion, BEVFormer v2 and StreamPETR.

Sparse4D algorithm architecture

##After three versions of iterations, the Horizon Sparse4D team It has successfully overcome technical problems such as improving the performance of sparse algorithms, reducing the computational complexity of timing fusion, and achieving end-to-end target tracking. Recently, they published a paper titled "Sparse4D v3: Advancing End-to-End 3D Detection and Tracking", detailing their research results. By using Horizon business data for performance verification, the Sparse4D team has successfully deployed on the Horizon Journey 5 computing solution. In the future, according to plans, Sparse4D technology will be used in Horizon’s next generation products. The achievement of this result will further promote the development of Horizon.

Dr. Yu Yinan, Vice President of Horizon and President of the Software Platform Product Line, pointed out that the current industry has entered the era of end-to-end sensing, which can be completed with only one network the entire perception task. He believes that the Sparse4D series of algorithms have improved the performance of sparse algorithms to a new level and successfully achieved end-to-end multi-target tracking. This is of great significance for both sparse perception and end-to-end autonomous driving, and can be said to be a milestone breakthrough. Horizon chose to open source Sparse4D to the entire industry, hoping to make progress together with outstanding developers in the industry.

Comparison between traditional perception system and end-to-end perception system

Another example of Horizon actively participating in the ecological construction of open source software for intelligent driving is the open source Sparse4D series of algorithms. This algorithm has huge application potential in the implementation of pure visual, end-to-end autonomous driving. In addition, Horizon has also open sourced leading technologies such as VAD algorithm and MapTR algorithm, which will further promote the development of the industry. It is expected that the Sparse4D algorithm will receive widespread attention and use by industry developers. Horizon's continued efforts will accelerate the industry's development process.

Horizon adheres to the concept of transforming independently innovative technologies, breakthrough products and solutions into the commercial value of ecological partners in the smart car industry, and contributes to the development of the industry. Energize. Through close collaboration, open integration, and cooperative innovation with all parties in the industry, Horizon is committed to becoming the source of living water for the smart automobile industry ecology and providing it with sustainable development momentum. Horizon fully understands that the mass production of autonomous driving is an important breakthrough, so we will continue to embrace open source and accelerate the implementation and mass production of cutting-edge technologies. We firmly believe that the future of win-win cooperation with the industry will be broader, and Horizon will continue to work hard to contribute to the prosperity of the smart car industry.

The above is the detailed content of Horizon releases open-source Sparse4D algorithm, pushing one step closer to end-to-end autonomous driving. For more information, please follow other related articles on the PHP Chinese website!