Robotics and sensing technology Team

Principal Investigator:Professor Jiun-In Guo

—

Summary

To assist Taiwan’s automotive industry to break through the barrier established by the existing ADAS patent portfolio in the world, we propose the idea of developing embedded AI deep learning technology for ADAS/niche self-driving applications so that Taiwan’s industry possess the chance to be part of the eco-system of ADAS/Self-driving products in the world. Our research focuses on three areas: (1) Automatic data labeling tool for deep learning applications and labeled ADAS datasets; (2) Embedded deep learning object detection/behavior analysis algorithm and model development; (3) Real-time deep learning computing platform development. Applications of our research include autonomous driving functions like lane keeping, automatic emergency braking (AEB), and automatic lane changing. In addition, we are also developing related technology required in certain niche self-driving applications, e.g., non-tracking AGV highly demanded in the Industry 4.0, to satisfy the needs of local industry.

Keywords

Embedded deep learning technology, fast data labeling, object detection and behavior recognition, ADAS, autonomous vehicle, industry 4.0

Innovations

• We have developed a world first fast labeling tool, ezLabel (shown in Fig. 1), to speed up 10x labeling efficiency as compared to the existing manual video labeling tools, which won two prizes in 2018 AUDI Innovation Award Taiwan.
• We have established a dataset with 11M+ of samples for deep learning object detection and behavior analysis for ADAS/Self-driving applications (as shown in Fig. 2). Out of the collected 11M+ samples, we have opened and shared 96K samples of them, most of which are related to vehicles, pedestrians, and cyclists under various weather conditions.
• We have developed an embedded deep learning model that can detect vehicles as far as 200m, which outperforms the Yolo v2 model (max. 50m) and owns over 10% higher in mAP as well.
• We have developed an embedded SSD lite model (SSD 512×512) that is suitable for using the TI TDA2X chipset for real-time object recognition. At 30fps, we obtained 72% mAP@Pascal VoC (as shown in Fig. 3).
• We have developed a deep learning technology based on 3D convolution neural network to predict whether rear vehicles will overtake in the next three seconds with a 95% or more accuracy (as shown in Fig. 4).
• We have developed a Taiwan first smart self-driving wheelchair for indoor mobility in places like airports and hospitals (as shown in Fig. 5).
• We have developed a world first bit accurate dynamic fixed-point deep learning model training and inferencing tool for CNN hardware accelerators.
• We have developed a hybrid fixed-point/binary CNN model training flow for CNN hardware accelerators (as shown in Fig. 6).

Benefits

• Have won two prizes in 2018 AUDI Innovation Award Taiwan.
• Have won an outstanding prize in AISlander 2018 with ezLabel and embedded AI technology.
• Have won a Technology Breakthrough Award in MOST 2018 Future Tech.
• Have won a Bronze award in 2018 MXIC design contest.
• Have conducted 18 cases of Industrial collaboration projects with total grant about NT$15,980K.