Robust, Multi-sensor and Deployable Hybrid SLAM

视觉里程计是一种在移动时仅用一个摄像头进行定位的能力。视觉SLAM(同步定位和地图绘制)将这一功能扩展到只使用一个相机同时绘制世界地图。其结果是一种地图和定位方法，只需一台廉价的相机就能生成类似LIDAR的地图。直接方法仅使用像素值，而间接方法使用检测到的视觉特征。混合方法利用直接方法和间接方法的优势。我们开发了一种混合方法，它利用了每种方法的优点，在计算效率较高的实施例中也克服了每种方法的缺点。大多数研究和商业SLAM系统都是间接方法，而直接方法只存在于研究实验室中。我们的方法具有亚像素精度、廉价的计算代价、在纹理剥夺下的健壮性等直接优势，并且同时提供稀疏和半密集重建密度。间接优势包括对大视距的稳健性、对优化种子的较低敏感度和对光线变化的稳健性。这些综合优势提供了支持地图重用的全球地图查询能力。该方法在视觉检测中提供了最先进的低误码性能，优于所有现有的混合、直接和间接方法。在关闭后的全满贯模式下，我们的累积误差基本上为零。该项目的目标是进一步开发这项技术，使其为工业做好准备。这包括增加动态光度和几何校准，并增加紧密集成其他传感器的功能，如激光雷达、其他相机、IMU、GPS等等。我们还打算通过开发一个健壮的位置检测和识别系统来改进环路闭合检测，该系统也可以用于选择关键帧。我们的目标是拥有一个可以使用多个摄像头的系统，并且可以轻松部署和健壮地用于各种自主地图应用，包括机器人、车辆、高清地图和其他应用。

Visual odometry is the ability to localize with only a camera while moving. Visual SLAM (Simultaneous Localization And Mapping) extends this to also mapping the world simultaneously with only a single camera. The result is a mapping and localization method that produces LIDAR like maps with only an inexpensive camera. Direct methods use only pixel values while indirect methods use detected visual features. Hybrid methods leverage the advantages of Direct and Indirect approaches. We have developed a hybrid approach that leverages the advantages of each method which also overcomes the shortcomings of each approach in a computational efficient embodiment. Most research and commercial SLAM systems are Indirect methods while Direct methods are only present in research labs. Our approach provides Direct advantages such as sub-pixel accuracy, inexpensive computational costs, robustness under texture deprivation and provides both sparse & semi-dense reconstruction densities. Indirect advantages capitalized include robustness to large view distances, less sensitivity to optimization seeds and light change robustness. These combined advantages provide a global map query capability that supports map re-use. The method provides state-of-the-art low error performance in visual odomdetry, outperforming all existing hybrid, direct & indirect methods. In full SLAM mode after closure, our accumulated error is essentially zero. The objective of this project is to further develop the technology so that it is industry ready. This includes adding on-the-fly photometric and geometric calibration and adding the capability to tightly integrate other sensors such as LIDAR, other cameras, IMU, GPS to name a few. We also intend to improve the loop closure detection by developing a robust place detection and recognition system that may also be used to select keyframes. The goal is to have a system that can use more than one camera and is easily deployable and robust for various autonomous mapping applications with robots, vehicles, HD mapping and other applications.