Hi, I’m M0rtzz, in Chinese my name is 徐梓航.

I am a student majoring in computer science at Zhengzhou University, and I have been studying computer knowledge since 2021, focusing on writing software to contribute to the open source world andshare knowledge and innovation.

I have amassed substantial practical experience in the ROS technology stack.

I won three national first prizes in the field of robotics in 2023, with two championships and one bronze medal.

In the 2023 RoboCup China tournament（2023RoboCup机器人世界杯中国赛）, to realize a general service robot that can respond to natural language commands, the large language model (LLM) is used as the core of the robot. By leveraging the LLM’s In-context Learning (ICL) capability, providing the LLM with some instructive examples allows it to autonomously deduce which basic functions are needed to execute complex instructions, and to arrange the execution order and logical hierarchy of these functions. Subsequently, it activates the pre-set functional modules in ROS in an orderly manner, thereby efficiently and accurately completing the natural language tasks given by the user. Our method achieved the highest score in the GPSR project of the 2023 RoboCup@Home-Open Platform.

To address the issue of building a high-quality dataset within a short race course, we first manually annotated a small dataset and trained initial weights. Subsequently, we innovatively used an AI engine-driven X-AnyLabeling for semi-automatic annotation of a large dataset, avoiding the tedious and inefficient manual annotation process. Ultimately, all data will be used for the training of YOLOv5.

Finally, The implementation of these strategies made us the first team in the history of the RoboCup@Home to achieve a full score in the robot vision project “What is That”, leading with an absolute advantage over teams from Tsinghua University, Southeast University, Shanghai University, and others, to win the championship..

In the 2023 China Robot Competition Special Competition General Purpose Service Robot Project（2023中国机器人大赛专项赛通用服务机器人赛项）, To achieve planar grasping of a robotic arm, a GGCNN (Generative Grasping CNN) model is constructed. It takes high-precision depth maps captured by RealSense as input and utilizes the end-to-end learning mechanism of CNN to output a pixel-level grasp pose map. This includes the optimal grasping point, grasping width, and rotation angle around the Z-axis. The model training enhances the authenticity of the dataset through Gaussian noise, thereby improving the grasping accuracy and robustness of the robotic arm in complex environments.

Finally, in the competition, I led the team to surpass teams from Xi’an Jiaotong University, Northwestern Polytechnical University, and others to win the championship with an absolute advantage.

• VisionVoyage
• An autonomous driving simulation system based on fish eye camera and other perception technologies has been implemented using UE4.

• os-design
• Modify the Linux-0.12 kernel source code by adding two system calls. The first system call copies the user state string content to the kernel state and saves it, while the second system call copies the saved string back from the kernel state to the user state. I have written an automated configuration script that automatically installs compilation toolchains such as gcc-3.4 and downloads the source code of bochs-2.2.5. Use the sed command modification configure script to add gcc compilation parameters to prevent compilation errors. Then, debug the modified Linux-0.12 kernel using the compiled bochs-x86 emulator. The Git submission style of this repo fully follows the Angular + gitmoji specificationn.

• m0rtzz.blog
• My website’s front-end source code.

• latex-resume-template
• Using LaTeX, a resume template was designed by writing the font of fontame-4.7.0 as a .sty style file and introducing it.

• linux-dynamic-wallpaper
• Through research, I have discovered a method of unpacking scene.pkg and generating .mp4 files.

FisheyeSegNet: Restricted Deformable Convolution based Semantic Segmentation Using Surround-View Fisheye Camera for Autonomous Driving

To address the issue of significant geometric distortion of targets in panoramic fisheye images, an end-to-end semantic segmentation network for panoramic fisheye images, FisheyeSegNet, has been constructed. The Multi-Scale Depth Aggregation Network (MSDAN), designed with the integration of constrained deformable convolution, effectively processes and represents multi-scale features, enhancing the network’s ability to model distortion. The proposed constrained deformable convolution and Flexible Attention Module (FAM) establish a dynamic receptive field, enhancing the model’s ability to adaptively capture distorted target features, increasing the model’s representational capacity, and guiding the network to correctly focus on the target objects.

To tackle the problem of data for fisheye vision research, the Cubemap algorithm was designed, and the M0rtzzWoodscape large-scale semantic segmentation dataset was constructed using CARLA. This dataset includes 8000 simulated images with precise pixel-level semantic annotations across 24 categories, compensating for the lack of large-scale training datasets.

Extensive experiments on public datasets such as Woodscape have verified the effectiveness and excellent generalization of this method. The experiments have been completed, and the paper is currently being written.

I excel in Linux, proficiently utilizing Shell to perform human-computer interaction.

I am familiar with C/C++, Python, Pytorch, Docker, Git, etc.

In addition to my daily backend development tasks, I can also handle basic website construction and deployment.

Furthermore, I have experience developing service robots using C/C++, Python and ROS.

I am keen on sharing knowledge as well as writing articles and tutorials regularly.

Previously, I primarily wrote in Chinese, but currently, my goal is to write articles in English.