Core content of the report:
"Unmanned" looks beautiful, but it has always been difficult to land
"Unmanned" is one of the earliest artificial intelligence application scenarios, Google, Apple, Tesla, Baidu and other domestic and foreign technology giants have been actively deployed since 2016, but until now it has been difficult to achieve large-scale commercial landing. We believe that 1) the high cost of obtaining and labeling multi-dimensional data, 2) there is still a big gap between the decision-making accuracy of small probability events and human beings, and 3) the unclear attribution of legal rights and responsibilities at the time of the accident is part of the problems restricting its development. We believe that the emergence of large models represented by ChatGPT and SAM will change the working paradigm of all industries, including intelligent driving, and we will initially explore some of these development opportunities from the perspectives of data, algorithms, and computing power.
Data: Large models improve the efficiency of data collection and data annotation
Massive data is the basis for intelligent driving/unmanned driving. Through the development of these years, an L2+ level smart electric vehicle can now usually collect multi-dimensional data of more than 10+ cameras, 1-2 lidars, and 3-5 millimeter-wave radars, and the data is labeled and used to train the model. The emergence of large models can first be done
1) Build virtual scenes to manually generate data to supplement the situation that is difficult to obtain/insufficient data in reality. The virtual simulation of Tesla FSD, NVIDIA's Omniverse are all representatives. 2) Data labeling is a very time-consuming and laborious work, and the emergence of a large image segmentation model represented by Meta's SAM can greatly reduce the cost of data annotation.
Algorithm: Large model improves perception accuracy, shadow model learns human driving habits
Intelligent driving algorithms mainly include 1) perception (recognizing roads and objects on the road), 2) prediction (predicting the behavior of surrounding vehicles and pedestrians), and 3) decision-making (controlling actions such as vehicle speed direction).
Major companies such as Tesla and New Forces began to adopt new technologies such as large models based on Transformer a few years ago, 1) improve the recognition accuracy of roads and objects; 2) learn human driving habits (shadow mode), 3) shorten the time required for decision-making, so that the model is more "anthropomorphic".
Intelligent driving industry chain: the trend of domestic substitution is significant
Benefiting from the objective demand for the localization of intelligent driving, the improvement of product performance of domestic suppliers in all links of the industrial chain and the rise of downstream independent brands, the trend of domestic substitution of parts and components is significant.
1) Chip: The gap between domestic players such as Horizon and Black Sesame and overseas manufacturers has gradually narrowed, and the localization service capability is stronger.
2) Domain controllers and solutions: Domestic players Desay SV, Jingwei Hengrun, Zongmu Technology, Zhixing Technology, etc. have been launched on a large scale, and the technological maturity has been continuously improved.
3) LiDAR: Domestic suppliers Hesai, Tudatong, Sagitar, etc. are in mass production at a faster pace.
4) 4D millimeter wave: domestic players include radar manufacturers such as Xingyidao and Senstec and chip companies such as Gatland (MMIC chip).
5) High-speed connector: Rosenberg has deep accumulation of technology, and electric connection technology, Ruida and other accelerated catch-up.
The following is an excerpt from the report, the original method of downloading the report is shown at the end of the article:
The full text of the report is shared in PDF and the data is obtained
Click the avatar to enter the homepage, follow this headline number, and click [My Services] to get the report