Haowen Lai

The project focuses on merging large-scale 3D point cloud maps without any prior information about their relative transformation. These large-scale point clouds, which are usually generated using SLAM with 3D LiDARs in the outdoors, differ a lot from the ones created by 3D scanners or RGB-D cameras. The noise becomes larger and the points are sparser, making most descriptors designed for point cloud registration fail. Furthermore, the issue of time consumption must be taken into consideration as the scale of maps becomes increasingly larger. To handle it, we propose a novel method that selects key points at a higher sematic level rather than choose them randomly as most current works do. Then features are extracted and descriptors are constructed by taking their neighbors into account. For details, please refer to our paper.

The abilities of self-localization and target perception for robots are of vital importance since they are the basic of other upper level functions, such as path planning, obstacle avoidance and navigation. Although some methods can provide the robot with localization results directly, they are somehow with weaknesses. For example, the use of external visual localization system may be unrealistic in practice, and GPS/RTK is rarely working in indoor enviroment. Thus, robots should be able to locate themselves with the sensors they carry. The capability of target perception is also signifiant in some tasks such as searching, game playing and autonomous driving. It is unreasonable to expect information sharing from the opponents in real scenes like most theory analyses do. In this project, we leverage 3D LiDAR and IMU to provide robots with the abilities of self-localization and target perception in pursuit-evasion games. The results show that our method is accurate and effective in both indoor and outdoor enviroment. For more details, please refer to our papers.

The research in this project focuses on robot sensing, understanding and interaction with the unknown environment by itself. Some key technologies are utilized, such as SLAM, multi-sensor fusion, robot navigation, self-exploration, object detection etc. What we want to do is making robots as intelligent as human in understanding its enviroment, especially in an unknown area with no human instruction. Specific tasks include exploring the enviroment autonomously and building a map, dynamic obstacles avoidance, self localization, object picking by robot arm, etc.

This project aims to build an automatic reading system for pointer meters. The system is comprehensive, including image capture device, core reading algorithm and reading management GUI. Unlike other reading systems, the core reading algorithm applied in ours was proposed by our team (see [pub]), which is able to process both uniform and non-uniform scale meters (meters whose scales are unevenly distributed) regardless of their shapes. This ability undoubtedly enlarges the range of meters in application. In some circumstances such as power plant, it is even critical because most of their meters are of non-uniform scales, e.g. power factor meter and large measurement Ampere meter. Images are captured by an self-designed embedded device and transmitted through wireless network. Besides, reading management GUI is also provided for monitoring, recording and controlling the whole system.

We presented a robot arm system for grabbing and classification in this project. As we know, for most control algorithms of robot arms, models need to be built before the arms can precisely move to a certain position. It is not an easy job especially when your robot arms are of different types. Even for the same type, it is hard to acquire an accurate model due to the manufacturing errors. So we use deep learning to make the arm learn the model by itself through practicing. The location of its joints acquired by the visual positioning module is used as training samples together with the control values of each motors. To Illustrate the adaptability and effectiveness of our approach, the algorithm is applied to two different robot arms for grabbing and classification tasks through collaboration. If the classification features are blocked by the clip of one arm when holding the object, the arm will pass it to another and let it observes the features.

This project focused on how to provide convenient UAV control and better user experience in selfie. In our research, the remote controller is discarded. Instead, users control the UAV for a certain task such as stopping following, resuming and landing through hand gestures. The gesture recognition module first segments hand area by color feature, then uses neural network to infer its meaning. The segmentation procedure dramatically reduces the time compared to using CNN to infer on a 3-channals color image directly, and requires less training samples. Another distinguishing characteristic is that our UAV utilizes a depth camera to locate the user and will automatically follow the person when moving, with no worry of getting lost. Other functions include automatic finding best selfie position, group shots and video recording, etc.