Application of artificial intelligence in TDC mining machine — automatic marking of traffic big data

The development and application of artificial intelligence technology

TDC mining equipment has been put into production now. This powerful equipment has a built-in artificial intelligence chip, which is mainly used for data annotation.

Since 1956, artificial intelligence has experienced 60 years of exploration. During this period, artificial intelligence has experienced six stages of development:

From 1956 to the early 1960s, it started to develop. The concept of artificial intelligence has been put forward, the machine theorem has been proved, and the checkers program has proved the feasibility of artificial intelligence.

From the 1960s to the early 1970s, we should reflect on the development stage. People gradually find that artificial intelligence is not omnipotent, its ability is very limited.

From the early 1970s to the middle 1980s, the application development stage. It is found that artificial intelligence can be used to solve problems in specific fields, so the research of artificial intelligence has shifted from general reasoning strategy to application of specialized knowledge.

From the middle of 1980s to the middle of 1990s, it was in a low development stage. Because of the defects of various technical bases, AI database lacks data base, single reasoning method and distributed function.

From the mid-1990s to 2010, the stage of steady development. The development of computer technology and Internet technology has brought a qualitative leap to artificial intelligence technology. In 1997, IBM’s dark blue computer defeated Kasparov, the chess champion. In 2008, IBM proposed the concept of intelligent earth.

From 2011 to now, the vigorous development stage: with the development of big data, cloud computing, Internet of things and other information technologies, ubiquitous sensing data and graphics processors and other computing platforms promote the rapid development of artificial intelligence technology represented by deep neural network, which has greatly crossed the “technology gap” between science and application, and realized the change of artificial intelligence technology from “useless and bad” The breakthrough of technology from “to” can be used “ushered in a new climax of explosive growth.

In 2016, Lee Shishi, a South Korean go player, lost to Google artificial intelligence robot after three hours of fierce battle. Artificial intelligence became famous in the first World War. At one time, the media and people’s discussion center became artificial intelligence. Since then, many activities can see the figure of artificial intelligence.

In 2017, openai five participated in the world-class competition ti8 of the large-scale MoBa game dota2. Openai five, which is composed of five neural networks, defeated a number of singles in the single solo mode, including the first single dendi at that time, and beat amateur teams in the subsequent competitions.

AI’s success in go and MoBa games means that AI can grasp the nature of chaos and continuity in the world through algorithms, which lays the foundation for the realization of artificial intelligence technologies such as image classification, speech recognition, Knowledge Q & A, human-computer game, unmanned driving, etc.

Automatic annotation of traffic data

AI can deal with complex problems such as image classification, speech recognition and just question and answer. The key lies in the adoption of convolutional neural network and deep learning technology. Traffic data automatic annotation is essentially image processing and image classification. The significance of data annotation is to help them understand and understand the world.

The objects of traffic big data annotation are mainly images, texts, videos, etc. the contents of image and video annotation usually include portrait, buildings, plants, roads, traffic signs, vehicles, etc. According to the analysis, the sensor group of an autonomous vehicle generates 10–20 TB of data every day. Such a large amount of data is analyzed and labeled by people, which is too expensive and inefficient.

For example, our common Imagenet data set was completed by nearly 50000 taggers from 167 countries around the world in two years. For some complex annotation scenarios, such as semantic segmentation, it takes an average of 19 minutes to annotate a picture on a coco dataset, and it takes such a long time to annotate a picture of several hundred K and several m, not to mention the dynamic 10–20tb traffic data.

The purpose of annotation is to let the machine understand and understand the world. It needs to let the machine understand what each element in the picture is, and distinguish each element. It needs to realize several key points — semantic segmentation, 2D / 3D frame line annotation, line segment annotation and fine regression.

Semantic segmentation — as shown in the above figure, the vehicle is marked in blue, the person is marked in red, the road sign is marked in yellow, and the sidewalk is marked in rose red.

2-D / 3-digit frame line marking — mark vehicles and pedestrians with frame lines as shown in the figure.

Segment annotation — annotates road markings.

Fine regression refers to the combination of semantic segmentation, 2D / 3D frame line annotation, line segment annotation, supplemented by interactive image segmentation technology, fine border technology, etc., to generate more accurate edge information. By using migration learning, it can be extended to a new untrained data set, further remove segmentation adhesion, approach the real image boundary, and generate refined traffic dynamic data.

Intelligent transportation is a very complex management system. The data types marked by TDC are far more than vehicle, pedestrian and road markings. TDC also requires artificial intelligence to mark the gravel, water, ice and other information on the way out surface. The shape and size of these things are not consistent, so it requires a high level of artificial intelligence in-depth learning.

In the cooperation between TDC and Huawei and JD, Huawei and JD said that the traditional annotation service of traffic data takes tens of minutes and costs more than 20 yuan for a single map, while the use of artificial intelligence automatic annotation can save more than 80% of the time, reduce the labor cost, avoid problems such as irregular data caused by artificial annotation, enhance the consistency of annotation, and improve the annotation Accuracy.

TDC has developed a relatively complete automatic annotation system, which has annotated tens of thousands of traffic photos and more than 300 Pb of traffic video. Facing the huge traffic scene, the next thing TDC needs to do is to promote the mining machine to be put into the market as soon as possible, and use crowdsourcing to enrich the data sources to solve the problem of data collection quality.

About TDC

TDC is an intelligent decision-making service platform based on transportation big data. Integrating the hardware equipment of the Internet of things, artificial intelligence and blockchain technology, it aims to reshape the traffic data specification, subvert the traditional data collection and cleaning mode, and output it to the demander in the form of standardized interface. Strive to fundamentally solve the problems of no real-time data and low data utilization in the transportation industry.