AI, which is the abbreviation of artificial intelligence and self-driving car are two popular research across the globe recently. Self-driving car has a significant application field in AI field. It is estimated that self-driving car’s related output value in the world will be 42 billion US dollars in 2025. Under the financial support of the Ministry of Science and Technology, the team of Professor Jiun-In, Guo, which is from the Institute of Electronics, National Chiao Tung University, combined AI with self-driving car and developed the technology of " The intelligent eye of Self-Driving - Embedded AI Object Detection System", has been cultivated for eight years. The research results have surpassed the international research status. It has gained favor with 28 domestic companies, and 59 collaborations between industry and college, technological transfer and technical consultation, winning abundant benefits from Industry and academia, since 2016.
The automatic labeling tool (ezLabel) won two awards from the AUDI Innovation Award
AI artificial intelligence is as smart as human beings. It can analyze and identify various objects in its images for humans. These works are attributed to human AI-labeled video teaching materials, complex software algorithms and powerful hardware computing capabilities.
Professor Guo’s team aims to achieve the self-driving/advanced driving assistance system (ADAS) by analyzing images. Successfully developed embedded computer vision deep learning technology. Generating a large numbers of databases for AI learning datasets through the rapid automatic labeling tool. Combining the team’s development, real-time software algorithms, conduce the cost for AI computer vision through computing platforms, which is requiring expensive GPU. Prof. Guo said that the team has developed the first set to label the video data fast by automatic labeling tool (ezLabel 2.0), which is used to label and prepare materials for AI learning. Its labeling efficiency is 10-15 times faster than the existing manual labeling data tools. ezLabel 2.0 won the two awards in Taiwan AUDI’s 1st Innovation Award, which was hosted by AUDI Auto. In addition, it has been tested by many domestic manufacturers and has made a quite good response. What’s more, Prof. Guo has built more than 15 million self-driving image databases for Taiwan, which will help developing AI self-driving object identification technology for Taiwan.
Embedded deep learning model detects vehicles up to 200 meters away, 4 times beyond existing technology
In the development of high-precision embedded AI depth learning algorithms, team of prof. Guo has developed an embedded deep learning algorithm that can detect vehicles up to 200 meters away, surpassing the current leading document algorithm (YOLO v2) four times, and its accuracy is higher than YOLO v2 about 10% mAP (average accuracy) under same computational complexity. It can perform real-time computing on Nvidia self-driving car platform (DRIVE-PX2) for all kinds of weather, Self-Driving car and ADAS applications. In addition, the team also developed the AI Deep Learning Object Recognition Algorithm (NCTU SSD lite) for low-power embedded SOC applications. The model’s size is only 7% of YOLO v2 with the same accuracy, and the computational complexity is only 27% of the YOLO v2, that can achieve instant object detection performance on TI TDA2X and car grade’s AI chip (iCatch V37) developed by domestic manufacturers. Furthermore, Professor GUO’s team combines the embedded object detection and semantic object segmentation to develop a multi-task ADAS system that is able to achieve multiple ADAS functions including LDWS, FCWS, RCWS, ACC, AEB, and BSD.
The first deep learning behavior prediction technology (to predict whether the vehicle will overtake in 3 seconds)
Team of Prof. Guo has extended the deep learning technology to predict the behavior of object and even control the driver, and has developed world’s first deep learning behavior prediction technology to predict whether the rear vehicle (car or scooter) overtakes. It can accurately predict the rear vehicle while driving. It can be used as the guardian for drivers’ safety. The potential of the embedded AI self-driving car rapid data labeling tool, self-driving car map, object detection and behavior prediction deep learning technology produced by Professor Guo's team is quite large. Just now, there are 28 cooperative manufacturers, and potential cooperation in the future, including AI chip companies, car power systems companies and self-driving car companies.
Department of Engineering and Technologies
Ministry of Science and Technology