Andy Lai is a passionate Computer Vision Engineer, a Deep Learning enthusiast, and a Master of Electronic and Computer Engineering. He loves Programming, Developing, and Problem Solving. Moreover, he is a highly proactive and fast learner who always makes good use of his free time to enhance his ability through online courses or books. Most importantly, he is always up for the challenges and willing to take difficult tasks!
The information in two-dimensional space (RGB image) is not enough to accurately describe the driving environment. Combined with three-dimensional information (Lidar), using Lidar and RGB information can complementing the advantages and disadvantages of different sensors, and realizing an object detection model in three-dimensional space.
Compared to the state-of-the-art 2D object detection networks which can achieve over 90% of accuracy, 3D object detection networks (only about 75% of accuracy) still have a lot to be improved. What I have done is to improve the execution speed and performance, such as using more compact feature representation of Lidar data to express features in three-dimensional space, using data augmentation and regularization methods to better deal with over-fitting problems, and using adaptive fusion methods for sensor information fusion.
Participated in the Industrial Technology Research Institute's project "Pre-crash Detection - Applied to Autopilot Systems". In this project, the model can detects the driving lanes and predicts the vehicle trajectory through a monocular camera (without depth sensor) to achieve pre-crash warning. Among them, using SSD for object detection, Mask RCNN for lane detection, LSTM for predicting vehicle trajectory ahead, and Self-supervised method estimated depth. Combine the above features altogether to make decisions and determine the possibility of impending collision.
For my work in this project is object detection part, I used a modified SSD to achieve faster execution time for real-time implementation.
Participated in the Industrial Technology Research Institute's project "Video Annotation Tool - Applied to Autopilot Systems". The main purpose of this project is annotate local video with object detection methods and use crowd-sourcing for verifying and correcting annotations. Then, we use the correct annotations to update the model through Weekly-supervised methods.
For my work in this project is the annotation part, which is combined object detection and object tracking methods to obtain highly reliable results.
In addition to the on-campus courses, I am very excited to be able to attend the "Machine Learning for Big Visual Data" course at IEEE International Elite School. Professor Huang from the Department of Electrical Engineering at the University of Washington, USA, teaches computer vision and image processing techniques, and introduces the state-of-the-art methods. Also, in-depth discussion of supervised learning of visual data, from neural networks, deep learning to the application of object tracking. It was an invaluable experience, which let me have different insights of understanding in this field, and also learn a lot of information and knowledge that is not available in the on-campus courses.
I am passionate of exploring new technologies and solving problems, there are full of resources on the Internet waiting me to explore. Except reading the latest paper I frequently watch online courses, such as Stanford University's CS231n: Convolutional Neural Networks for Visual Recognition and CS224d: Natural Language Processing with deep Learning, University of Toronto's CSC321: Intro to Neural Networks and Machine Learning, etc.
National Taiwan University of Science and Technology, Taipei, Taiwan
Sep. 2016 - Jun. 2018
National Taiwan Ocean University, Keelung, Taiwan
Sep. 2012 - Jun. 2016
Andy Lai is a passionate Computer Vision Engineer, a Deep Learning enthusiast, and a Master of Electronic and Computer Engineering. He loves Programming, Developing, and Problem Solving. Moreover, he is a highly proactive and fast learner who always makes good use of his free time to enhance his ability through online courses or books. Most importantly, he is always up for the challenges and willing to take difficult tasks!
The information in two-dimensional space (RGB image) is not enough to accurately describe the driving environment. Combined with three-dimensional information (Lidar), using Lidar and RGB information can complementing the advantages and disadvantages of different sensors, and realizing an object detection model in three-dimensional space.
Compared to the state-of-the-art 2D object detection networks which can achieve over 90% of accuracy, 3D object detection networks (only about 75% of accuracy) still have a lot to be improved. What I have done is to improve the execution speed and performance, such as using more compact feature representation of Lidar data to express features in three-dimensional space, using data augmentation and regularization methods to better deal with over-fitting problems, and using adaptive fusion methods for sensor information fusion.
Participated in the Industrial Technology Research Institute's project "Pre-crash Detection - Applied to Autopilot Systems". In this project, the model can detects the driving lanes and predicts the vehicle trajectory through a monocular camera (without depth sensor) to achieve pre-crash warning. Among them, using SSD for object detection, Mask RCNN for lane detection, LSTM for predicting vehicle trajectory ahead, and Self-supervised method estimated depth. Combine the above features altogether to make decisions and determine the possibility of impending collision.
For my work in this project is object detection part, I used a modified SSD to achieve faster execution time for real-time implementation.
Participated in the Industrial Technology Research Institute's project "Video Annotation Tool - Applied to Autopilot Systems". The main purpose of this project is annotate local video with object detection methods and use crowd-sourcing for verifying and correcting annotations. Then, we use the correct annotations to update the model through Weekly-supervised methods.
For my work in this project is the annotation part, which is combined object detection and object tracking methods to obtain highly reliable results.
In addition to the on-campus courses, I am very excited to be able to attend the "Machine Learning for Big Visual Data" course at IEEE International Elite School. Professor Huang from the Department of Electrical Engineering at the University of Washington, USA, teaches computer vision and image processing techniques, and introduces the state-of-the-art methods. Also, in-depth discussion of supervised learning of visual data, from neural networks, deep learning to the application of object tracking. It was an invaluable experience, which let me have different insights of understanding in this field, and also learn a lot of information and knowledge that is not available in the on-campus courses.
I am passionate of exploring new technologies and solving problems, there are full of resources on the Internet waiting me to explore. Except reading the latest paper I frequently watch online courses, such as Stanford University's CS231n: Convolutional Neural Networks for Visual Recognition and CS224d: Natural Language Processing with deep Learning, University of Toronto's CSC321: Intro to Neural Networks and Machine Learning, etc.
National Taiwan University of Science and Technology, Taipei, Taiwan
Sep. 2016 - Jun. 2018
National Taiwan Ocean University, Keelung, Taiwan
Sep. 2012 - Jun. 2016