VINS MONO === ###### tags: `暑期集訓` [TOC] # Sensor Data ## IMU Measurement - bias 誤差 - noise 雜訊 ## Camera Measurement ### pinhole model: 顯示的大小與真實大小呈現一個倍率,以f~x~ f~y~ 表示 ![](https://i.imgur.com/3izhkli.jpg =500x) 校正出K ![](https://i.imgur.com/KaoSpGf.jpg =500x) ### challange - Scale 不確定(只能確定移動方向,不知道移動距離) - **與初始化有關** - Up-to-scale motion estimation and 3D reconstruction - **與optimization有關** ## Vision Data process ### KLT tracker(Kanade-Lucas-Tomasi Tracking Method) - 光流法 ### RANSAC(隨機抽樣一致算法 - RANdom SAmple Consensus,RANSAC - 驗證 ### Keyframe 1. Case1: 可以分辨出旋轉與平移 2. Case2: 要選擇好的Key frame - 變化不大不用多做一個key frame ### Notation - 外部參數 imu to camera frame - 從body frame 轉到 world frame # System framework ![](https://i.imgur.com/MLcnPs1.jpg) # initialization :::warning Very Very Important ::: 1. 設定外部參數 2. IMU 校正 (acc and gyrop 的 biases) ![](https://i.imgur.com/h7Dz3xJ.jpg) # Estimation Process # Visaul-Inertial Nonlinear Optimization ## 深度不準的問題 ## cost function ![](https://i.imgur.com/7IPNjOH.jpg =500x) ![](https://i.imgur.com/Q6YCDkb.jpg =500x) 算出同一個特徵點由不同frame 轉換後比較的殘差 ![](https://i.imgur.com/OyLPvmb.jpg) 迭代求解 ![](https://i.imgur.com/XRBqQ1G.jpg) # Marginalization 高斯牛頓 邊緣化 # monocular vision-based https://blog.csdn.net/u010821666/article/details/52915238 - ==VIO== - PTAM - SVO - LSD-SLAM - DSO - ORB-SLAM popular EKF based VIO: MSCKF