解读DP-SLAM （1）

解读DP-SLAM （1）

解读DP-SLAM分为两部分：第一部分是解读论文，第二部分是解读代码。

#1  从Abstract中可以得到一些信息：

(1) a novel laser range finder based algorithm for SLAM

(2) no predetermined landmarks; purely laser based

(3) uses a particle filter to represent

      a. robot pose

      b.possible map configurations ; called distributed particle (DP) mapping

从Introduction可以得到一些信息：

(1) Even with an accurate laser range finder, map-making presents a difficult challenge:

      a. a precision position estimate is required to make consistent updates to the map

      b. producing accurate maps in real time, based on a single pass over the sensor data, without an

           of line correction phase

(2) A straight forward approach:

      localize the robot based upon a particle map and then update the map based upon the maximum

      likelihood position of the robot      (but errors accumulate over time)

(3) EM algorithm:

      it involves an expensive off-line alignment

(4) FastSLAM:

      a. does not require explicit loop-closing heuristics

      b. using a Rao-Blackwellized particle filter to sample robot poses

      c. track the position of a fixed number of predetermined landmarks using a Kalman filter

(5) We present a method:

      a. purely laser based an makes no landmark assumptions

      b. avoid the data association problem by storing multiple detailed maps instead of sparse landmarks

      c. uses a particle filter to represent robots poses and map configuration

      d. no assumptions about the environment

      e. DP-SLAM could bu further complimented by inclusion of existing algorithms for closing loops

#2     进入文章的第二部分，Particle Filters for Localization and Mapping

#2.1 简介粒子滤波：

Particles: a weighted and normalized set of sampled states, S = {s_1,..,s_m}

滤波过程：

At each, upon observing an observation "o"

1. Samples m new states S' = {s'_1, ..., s'_m} from S with replacement

2. Propagates each new state through a Markovian transition model: P( S''| S')

3. Weights each new state according to a Markovian observation model: P(o|S'')

4. Normalizes the weights for the new set of states

#2.2  简介Particle Filters for Localization (只考虑定位假设地图是已知的)：

粒子滤波的基本结构(all of which are noisy)：

a. The robot pose is the hidden state to be tracked

b. The state transition is the robot's movement

c. The observation are the robot's sensor readings

小车在2D空间运动模型：

小车在O-XY平面运动，考虑它沿X轴的位移以及速度，沿Y轴的位移及速度，以及小车的转向角度及角速度（相对于X轴的角度）

通过观测数据给粒子加权的方法 (weight the particles based on the robot's current observation)：

下次再来讨论