How can self-driving vehicles use spatio-temporal information to better identify moving objects, and how can they know "where I am" without positioning and navigation? This is a technical problem that self-driving vehicles are currently overcoming. Now, Chinese self-driving AI company Fei Mo Zhixing has come up with a better solution.
html On June 30, the two latest research results of Hao Mo Zhixing were successfully selected for IROS (IEEE/RSJ International Conference on Intelligent Robots and Systems International Conference on Intelligent Robots and Systems) 2022, the top academic conference in the field of robotics. They will be published in the IROS 2022 conference in the near future. published. The two latest paper research results submitted by the Hao Mo Zhixing team are: "Efficient Spatial-Temporal Information Fusion for LiDAR-Based 3D Moving Object Segmentation" ("Efficient Spatial-Temporal Information Fusion for LiDAR-Based 3D Moving Object Segmentation") , "OverlapTransformer: An Efficient and Yaw-Angle-Invariant Transformer Network for LiDAR-Based Place Recognition" ("OverlapTransformer: An efficient, rotation-independent position recognition network based on LiDAR"). Two papers stood out from more than 2,000 submitted papers and were successfully selected. Combining the application of lidar in autonomous vehicles, the paper proposes a new deep neural network for lidar moving target segmentation and a new lidar-based position recognition algorithm to help autonomous vehicles effectively utilize It uses spatio-temporal information, identifies moving targets, and quickly and accurately positions itself, thereby greatly improving the perception capabilities of lidar.
With the gradual and in-depth exploration of lidar applications in the field of autonomous driving in recent years, its powerful spatial three-dimensional resolution capability has been generally regarded as an important capability in the process of upgrading autonomous driving technology to high-end and commercialization. However, the information data collected by the hardware also requires faster and more accurate analysis by algorithms to help autonomous vehicles make better use of it and achieve safer driving. Hao Mo Zhixing's two papers start from this perspective.
pointed out in "Efficient Spatial-Temporal Information Fusion for LiDAR-Based 3D Moving Object Segmentation" that accurate moving target segmentation is an important task for autonomous driving, and how to effectively use spatiotemporal information is the key to 3D LiDAR moving target segmentation. The key issue. To this end, Hao Mo proposed a new deep neural network that uses spatiotemporal information and different representation modes of lidar to improve lidar MOS performance. Specifically, Hao Mo proposed a novel and effective online moving target segmentation network based on lidar, which uses a dual-branch structure to better integrate spatial information and temporal information, and introduces a "coarse-to-fine" Strategies are used to reduce the boundary blur problem on the object boundary, and while maintaining real-time performance, the performance surpasses the previous network in one fell swoop. Currently, related methods achieve state-of-the-art lidar MOS performance on the SemanticKITTI MOS benchmark.
In the article "Overlap Transformer: An Efficient and Yaw-Angle-Invariant Transformer Network for LiDAR-Based Place Recognition", Hao Mo proposed a new position recognition method, using the three-dimensional lidar installed on the autonomous vehicle to generate Range images can be used to detect SLAM loop closure candidates or directly identify locations using only lidar data without using any other information, and can be well generalized to different environments without fine-tuning. Long-term location recognition is achieved in long-term outdoor large-scale environments (millimetre data set). OverlapTransformer runs faster than most state-of-the-art methods, and all indicators have reached SOTA.
IROS (IEEE/RSJ International Conference on Intelligent Robots and Systems, IEEE/RSJ International Conference on Intelligent Robots and Systems) is one of the most famous and influential top academic conferences in the field of robots and intelligent systems in the world. The 2022 IROS, with the theme of "Embodied Artificial Intelligence for a Symbiotic Society", will be held in Kyoto, Japan from October 23-27.
As the number one mass-produced autonomous driving company in China, Haomo Zhixing has gained wide recognition from both the industry and outside the industry in the past two and a half years with its strong technological innovation capabilities and rapid product launch capabilities.MANA, the first data intelligence system in China's autonomous driving field created by Haimo, improves the capabilities of autonomous driving products by defining and using data intelligence. It is the cornerstone of Haimo's product iteration and has the ability to mine data value with high efficiency and low cost. As of June 2022, the learning time of the MANA data intelligence system has exceeded 240,000 hours, and the virtual driving experience is equivalent to 20,000 years of driving by a human driver. Based on MANA's powerful capabilities, large-scale mass production capabilities and increasingly mature business model, Haimo has established a complete data closed loop, providing strong support for the continuous upgrading of its own technology products and the advancement of China's autonomous driving technology.
