Beijing (Gasgoo)- On January 22, Horizon Robotics made a significant move by open-sourcing its leading Sparse4D series pure vision autonomous driving algorithms. This initiative aims to foster collaboration among developers and encourage collective exploration into cutting-edge technologies like end-to-end autonomous driving and sparse perception.
The Sparse4D algorithm is now available on the GitHub platform, allowing developers to access the source code by following Horizon Robotics' official GitHub account, "Horizon Robotics."
Sparse4D represents a series of algorithms advancing towards long-sequence sparse 3D target detection, falling within the realm of temporal multi-view fusion perception technology. Aligned with the industry trend towards sparse perception, Sparse4D constructs a purely sparse fusion perception framework, enhancing algorithm efficiency and accuracy while simplifying the perception system.
Compared to dense BEV algorithms, Sparse4D reduces computational complexity, overcoming limitations on perception range imposed by computing power and surpassing dense BEV algorithms in both perception effectiveness and inference speed, said Horizon Robotics.
Through three iterations, Horizon Robotics’ Sparse4D team successfully addressed challenges in improving sparse algorithm performance, reducing temporal fusion computational complexity, and achieving end-to-end target tracking. Their recent paper, "Sparse4D v3: Advancing End-to-End 3D Detection and Tracking," details these advancements. With performance validation using Horizon's business data, Sparse4D is now deployed on the company’s Journey 5 computing platform. The next phase involves integrating Sparse4D into Horizon Robotics’ next-generation products, as per their strategic plan.
