SJTU JiaoLong RM2018 Armor Detection

环境要求

• Ubuntu 14.04 或更高
• CMake
• OpenCV（推荐OpenCV3以上版本）
• libMVSDK.so（全局曝光相机驱动）

编译运行

Release模式

mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make
./FindArmor

Debug模式

mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Debug ..
make
./FindArmor

状态机

• Fast Explore（默认状态）：在机器人快速移动的时候，寻找目标，有一定概率误识别
• Fast Track：追踪在Fast Explore中找到的可能装甲板
• Slow Explore：机器人慢下后，重新寻找目标，此时画面稳定，不容易误识别
• Slow Track：追踪在Slow Explore中找到的可能装甲板

寻找算法

• 筛选出高亮区域（对于黑白摄像头）或红/蓝区域（对于彩色摄像头）
• 使用findContour寻找轮廓
• 寻找长条形的轮廓，可能是两侧的灯条
• 配对灯条（根据长度、角度等）

追踪算法

• KCFTracker：传统方法的追踪器，可以达到速度（60fps）和性能的平衡

加速算法

• OpenMP：并行计算，在处理当前图像的同时，获取下一张图像

可改进

• 使用机器学习配对灯条
• 更进一步，直接使用目标检测查找装甲板
• 识别中间贴纸的数字
• 同时使用两个摄像头

代码格式

Google Style

PS

• 我们使用了MindVision的全局曝光相机，驱动是专有的。在include/camera_wrapper.h里对驱动进行了包装。
• 为了提高速度，没有使用OpenCV自带的KCFTracker，另外找了一个CPP版本，被包装在src/KCFTracker.cpp里。
• 为了便于区分自己电脑的开发环境和妙算上的部署环境，在include/platform.h里设置了宏，通过是否ARM架构来区分。
• main.cpp里面还有一些宏来控制OpenMP，显示中间图像和录像。其中OpenMP和显示中间图像是互斥的。默认Debug模式下显示中间图像，关闭OpenMP，Release模式下不显示中间图像，开启OpenMP。
• 为了优化彩色相机的速度，直接处理了原始的拜耳阵列，可以通过include/bayer_hack.h中的宏BAYER_HACKING控制。涉及到的地方有相机驱动（是否传输原始图像）和Armor类（参数改变）。

---

Requirements

• Ubuntu 14.04 or higher
• CMake
• OpenCV (version 3 recommended)
• libMVSDK.so (Global Shutter Camera Driver)

Compile and Run

Release Mode

mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make
./FindArmor

Debug Mode

mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Debug ..
make
./FindArmor

State Machine

• Fast Explore(default state): Explore the armor when the robot is moving fast
• Fast Track: Track the armor detected in Fast Explore
• Slow Explore: After the robot slows down, explore the armor again
• Slow Track: Track the armor detected in Slow Explore

Explore Algorithm

• Find the light region(for Gray Camera) or the blue/red region(for Color Camera)
• Use findContour to find the contours
• Find the thin and long contours, which may be the side light of armor
• Pair the lights to find the armor(length, angle and so on)

Track Algorithm

• KCFTracker: the balance of speed(60 fps) and accuracy.

Speedup Algorithm

• OpenMP: Process the image and fetch next image at the same time.

Ways to improve

• Use Machine Learning to pair the lights
• Use Object Detection to explore the armor
• Recognize the digit in the center
• Use two camera at the same time

Code Style

Google Style

PS

• We use a special camera with Global Shutter, so the driver is special. It is wrapped in include/GlobalCamera.h.

• We find a kcftracker of cpp version instead of the one in the OpenCV, because it runs much faster. It is wrapped in src/kcftracker.cpp.

• To distinguish with the PC platform and MiniPC platform, there is a macro in the precom.h. If the CPU is ARM architecture, then it is regarded as the MiniPC platform.

• There are some macros in the precom.h to switch the OpenMP, Show Image and Record Videos. And when the OpenMP is on, there is no way to show the image.

• To improve the performance of RGB camera, we process the raw data with BAYER format. To learning more, search the Bayer. It is opened by macro BAYER_HACKING.