Current Issues of Deep Learning for Automated Driving
| 讲座名称 | Current Issues of Deep Learning for Automated Driving |
| 讲座时间 | 2018-04-04 09:00:00 |
| 讲座地点 | 北校区主楼I区313 |
| 讲座人 | Prof. Xinming Huang |
| 讲座人介绍 |  Xinming Huang is a Professor in the Department of Electrical and Computer Engineering at the Worcester Polytechnic Institute (WPI) and Director of the Embedded Computing and Intelligence Lab. WPI is located in the suburb of Boston. It was established in 1865 and was the third oldest engineering universities in the US. The university is ranked 60 among all national research universities by US News & World Report in 2017. Dr. Huang received his PhD in electrical engineering from Virginia Tech in 2001. After that, he joined the Bell Labs of Lucent Technologies as a Member of Technical Staff with the wireless advanced technology laboratory. He was a recipient of the Central Bell Labs annual excellence award, IBM faculty, DARPA young faculty award, IEEE HKN outstanding professor award, and WPI faculty achievement award. His main interests are in the areas of hardware architecture for wireless communications, error correction coding, information security, computer vision and deep learning. He is also an expert on cyber-physical systems for autonomous vehicles and smart health. He has over 100 publications on IEEE transactions and top conferences. |
| 讲座内容 | Automated driving has experienced a major setback owing to the Uber accident in Arizona. In this talk, we will discuss some of the current issues and challenges in deep learning particularly related to autonomous vehicles. Although the current deep learning algorithms have achieved many successes in image recognition and objection detection, i.e. ImageNet, there remains several major issues including accuracy, data, and speed. The performance of camera-based solution is subjected to light conditions and background. The accuracy of a pre-trained network cannot guarantee when applying to the practical scenarios. Multi-sensor fusion is a possible approach towards better accuracy and reliability. As an example, pedestrian detection using both thermal and stereovision shows significant performance improvement. Secondly, deep learning requires a large amount of data. It is laborious to collect and label new data continuously. Generating virtual data through simulation and augmentation can effectively improve the training of a neural network. Finally, we need to accelerate the processing speed of the deep learning neural networks. GPU is a popular platform for training neural network. However, FPGA and ASIC can be customized to provide very high throughput and low latency through novel network architecture and hardware implementations. Our research goal is to develop the key technologies for automated driving and demonstrate on our prototype vehicle. We actively seek research collaborations with industry partners to solve important research problems towards fully automated driving. |
转载请注明出处:西安电子科技大学学术信息网
如果您有学术信息或学术动态,欢迎投稿。我们将在第一时间确认并收录，投稿邮箱: meeting@xidian.edu.cn
如果您有学术信息或学术动态,欢迎投稿。我们将在第一时间确认并收录，投稿邮箱: meeting@xidian.edu.cn