Robotics-Robot Sensing (Gaussian Model)

Robot perception is the last course in the UPNN Robot Project, which uses visual methods to perceive the environment. Unlike the previously mentioned robot vision, robot perception is more focused on the identification and detection of environmental objects. Unlike computer vision, objects recognized by robot vision often do not require high-precision measurements, and objects have obvious characteristics. The most typical application of robot perception is the perceptual--slam of the environment, simultaneous positioning and map building. If robot vision solves the problem of where am I, then robotic Perception is it.

1, 1D Gaussian

Perception to solve is the problem of environmental recognition, along with the idea of PGM down, recognition is the probability of calculation. For general recognition tasks, such as identifying a tennis ball in a natural environment, you can model the color of the tennis ball. In order to meet people's perception, the RGB image can be converted to HSI image, where chroma hue, and brightness decoupling, become an invariant. It is only related to the material and lighting conditions. If the color of an object is uniform, its H will satisfy the Gaussian distribution. It is only necessary to extract this distribution from the training set, which can be used to determine whether the object exists in the scene.

2, Mutiple Gaussian

The mutiple Gaussian corresponds to a multivariable Gaussian model. The multivariate Gaussian model is divided into correlation and irrelevant two kinds, which are manifested on the covariance matrix. Its distribution mean is the sample mean, whose variance is the covariance matrix of the sample!

3, Expecatation maxiumazition

EM algorithm is one of the most important algorithms in unsupervised learning. It can be decomposed into two steps--e step, M step

Where E step is represented by the label---> Parameters. The label is not necessarily a 01 sample, it can also be the probability of each sample. If it is a probability sample, you can use weighted Gaussian estimation to estimate parameters: The specific algorithm is shown in PGM Week9 homework in Oschina.

M step is represented by the parametes---> label. This is the probability that the current parameter is calculated for each label of the sample.

In this way, stable classification results can be obtained in the end. It is important to note that the EM algorithm is very sensitive to the initial label. If there are a few missing marks in the equivalent sample, the EM algorithm can satisfy the demand well. If pure clustering, you can consider using other clustering algorithms to give the cluster results, and then use EM to optimize.

　　

　　

Robotics-Robot Sensing (Gaussian Model)