CN113763263A - Water mist tail gas noise treatment method based on point cloud tail gas filtering technology - Google Patents

Abstract

A water mist tail gas noise treatment method based on a point cloud tail gas filtering technology. In the prior art, training data is added, more data containing water mist tail gas are used for training a detection model or a classification model, a large amount of new data is needed, the generalization capability is poor, and the noise processing method can be easily interfered by the water mist tail gas and the like again when a new vehicle type or different scenes are met. The method comprises the following steps: firstly, after a CNN model is used for detecting obstacles in a point cloud, a plurality of obstacles in the point cloud picture can be obtained, the purpose of removing water mist noise can be achieved only by identifying and deleting water mist obstacles in the point cloud picture, and whether each obstacle is water mist can be known by judging as follows; (1) for one obstacle represented by a 3d rectangular box, rasterizing it in the xoy plane; (2) the frequency falling on the ith order is denoted by pi. The invention relates to a water mist tail gas noise processing method based on a point cloud tail gas filtering technology.

Description

Water mist tail gas noise treatment method based on point cloud tail gas filtering technology

Technical Field

The invention relates to a water mist tail gas noise treatment method based on a point cloud tail gas filtering technology.

Background

In an automatic driving scene of a mine truck, when the laser radar is used as a sensor to sense surrounding obstacles, the influence of tail gas of a vehicle is easy to be caused, especially when a truck has high power and a working environment is cold, the tail gas or water vapor generated by the vehicle is obvious, the interference is generated on the laser radar, the obstacle is easy to be detected mistakenly, and the main vehicle is subjected to frequent emergency braking and even braking;

in the water mist tail gas noise processing method in the prior art, a special classification model or a detection model is generally trained to detect water mist tail gas noise which may be encountered, or other sensor information such as camera information is combined to preprocess point cloud data, before a sensing system senses the point cloud data, the water mist tail gas noise is removed, a common noise solution is to increase the training data and use more data containing water mist tail gas to train the detection model or the classification model, but the mode needs a large amount of new data and has poor generalization capability, when an automatic driving system encounters a new vehicle type or different scenes, the automatic driving system is easily interfered by the water mist tail gas and the like again, and the scheme consumes computing resources and has higher computing time consumption; the other method is to project the image segmentation result onto the point cloud by combining the camera information and the laser radar, but the method has the disadvantages of obvious defects, poor generalization, more calculation resources and high possibility of being influenced by the light and shade of illumination.

Disclosure of Invention

The invention aims to provide a water mist tail gas noise processing method based on a point cloud tail gas filtering technology, which overcomes the defects that the traditional technology is easy to cause obstacle false detection, causes main car frequent and sudden braking and even braking and needs more computing resources, and effectively ensures the driving safety of an automatic mine truck driving car.

The above purpose is realized by the following technical scheme:

a water mist tail gas noise processing method based on a point cloud tail gas filtering technology comprises the following steps: firstly, after a CNN model is used for detecting obstacles in a point cloud, a plurality of obstacles in the point cloud picture can be obtained, the purpose of removing water mist noise can be achieved only by identifying and deleting water mist obstacles in the point cloud picture, and whether each obstacle is water mist can be known by judging as follows;

(1) for an obstacle represented by a 3d rectangular box, rasterizing the obstacle in the xoy plane, for example, dividing the obstacle into n × n rectangular grids, then calculating the number of obstacle point clouds in each grid, so as to obtain a matrix, wherein each number on the matrix represents the number of points in the grid at the position, the maximum number in the grid is max, the minimum value is min, setting a hyperparameter k, step = (max-min)/k,

thus, k levels can be obtained, and then the number of points falling in each level is counted, for example, the first level represents the range [ min, min + step ], …, and the last level represents [ min + (k-1) step, max ];

(2) denotes with pi the frequency falling on the ith order, then

The topological distribution of the point cloud can be represented, wherein the probability H is the larger the higher the probability H is, the probability of water mist is, and the probability of water mist is the smaller the probability H is;

(3) the more uniformly distributed obstacles between the points are the higher probability of water mist, and the less uniformly distributed obstacles are the water mist.

Has the advantages that:

1. the invention mainly provides a water mist tail gas noise treatment method based on a point cloud tail gas filtering technology, which can effectively overcome the defects that the prior art needs a large amount of new data, has poor generalization capability and is easily influenced by illumination brightness.

The specific implementation mode is as follows:

example 1:

a water mist tail gas noise processing method based on a point cloud tail gas filtering technology comprises the following steps: firstly, after a CNN model is used for detecting obstacles in a point cloud, a plurality of obstacles in the point cloud picture can be obtained, the purpose of removing water mist noise can be achieved only by identifying and deleting water mist obstacles in the point cloud picture, and whether each obstacle is water mist can be known by judging as follows;

(1) for an obstacle represented by a 3d rectangular box, rasterizing the obstacle in the xoy plane, for example, dividing the obstacle into n × n rectangular grids, then calculating the number of obstacle point clouds in each grid, so as to obtain a matrix, wherein each number on the matrix represents the number of points in the grid at the position, the maximum number in the grid is max, the minimum value is min, setting a hyperparameter k, step = (max-min)/k,

thus, k levels can be obtained, and then the number of points falling in each level is counted, for example, the first level represents the range [ min, min + step ], …, and the last level represents [ min + (k-1) step, max ];

(2) denotes with pi the frequency falling on the ith order, then

The topological distribution of the point cloud can be represented, wherein H is probability, the larger H is, the larger the probability of water mist is, and the smaller H is, the smaller the probability of water mist is;

(3) the more uniformly distributed obstacles between the points are the higher probability of water mist, and the less uniformly distributed obstacles are the water mist.

In the point cloud data, the water mist tail gas noise point cloud distribution and the common barrier point cloud distribution have very obvious difference, and the difference can be used for well distinguishing the water mist tail gas noise point cloud distribution and the common barrier point cloud distribution to achieve the purpose of denoising;

the water mist tail gas noise (hereinafter referred to as water mist) is distributed uniformly and can penetrate through, when the laser radar scans the water mist, a barrier with uniformly distributed points can be generated in the space, and due to the reasons of non-penetrability and self-shielding of the barrier in general meaning in road driving, the generated point cloud shape is usually one surface of the barrier which can be scanned by the laser radar, namely the point is concentrated on the surface of one side of the barrier, such as the point cloud shape of most vehicles, which is usually in an L shape. Through two point cloud topological structures with obvious differences, the two point cloud topological structures can be distinguished, the more uniform obstacles are distributed between the points, the higher the probability of water mist is, the obstacles with point distribution concentrated on one plane are, and the lower the probability of water mist is.

Claims (1)

1. A water mist tail gas noise processing method based on a point cloud tail gas filtering technology is characterized by comprising the following steps: the method comprises the following steps: firstly, after a CNN model is used for detecting obstacles in a point cloud, a plurality of obstacles in the point cloud picture can be obtained, the purpose of removing water mist noise can be achieved only by identifying and deleting water mist obstacles in the point cloud picture, and whether each obstacle is water mist can be known by judging as follows;

(1) for an obstacle represented by a 3d rectangular box, rasterizing the obstacle in the xoy plane, for example, dividing the obstacle into n × n rectangular grids, then calculating the number of obstacle point clouds in each grid, so as to obtain a matrix, wherein each number on the matrix represents the number of points in the grid at the position, the maximum number in the grid is max, the minimum value is min, setting a hyperparameter k, step = (max-min)/k,

thus, k levels can be obtained, and then the number of points falling in each level is counted, for example, the first level represents the range [ min, min + step ], …, and the last level represents [ min + (k-1) step, max ];

(2) denotes with pi the frequency falling on the ith order, then

The topological distribution of the point cloud can be represented, wherein H is probability, the larger H is, the larger the probability of water mist is, and the smaller H is, the smaller the probability of water mist is;

(3) the more uniformly distributed obstacles between the points are the higher probability of water mist, and the less uniformly distributed obstacles are the water mist.