CN-115424041-B - Sparse event point-oriented space-time clustering small target detection method

CN115424041BCN 115424041 BCN115424041 BCN 115424041BCN-115424041-B

Abstract

The invention discloses a sparse event point-oriented space-time clustering small target detection method, which comprises the steps of preprocessing original data shot by an event camera to obtain preprocessed event data, traversing all event points in the event data, endowing each event point with a weight value, arranging all event points in a descending order according to the size of the weight value, taking the average weight of M% of event points after sorting as a clustering threshold value, carrying out nearest neighbor clustering on the event data according to the clustering threshold value to obtain a preliminary detection result, and carrying out point cloud filtering on the preliminary detection result to obtain a final detection result. The invention can directly extract small target features from event data for detection, can effectively filter event point clutter triggered by thermal noise and interference event points generated by large object movement under a static background, and has high detection precision.

Inventors

LI MIAO
AN WEI
SHENG WEIDONG
LIN ZAIPING
ZENG YAOYUAN
Deng xinpu
AN CHENGJIN
Sun Zhezheng
WANG LONGGUANG

Assignees

中国人民解放军国防科技大学
中国人民解放军国防科技大学

Dates

Publication Date: 20260421
Application Date: 20220907
Priority Date: 20220907

Claims (9)

1. A sparse event point-oriented space-time clustering small target detection method is characterized by comprising the following steps of: preprocessing the original data shot by the event camera to obtain preprocessed event data; traversing all event points in the event data, giving a weight value to each event point, and arranging all event points in descending order according to the weight value; taking the average weight of M% of event points after sequencing as a clustering threshold value, and carrying out nearest neighbor clustering on the event data according to the clustering threshold value to obtain a preliminary detection result; the primary detection result obtaining process comprises the following steps: Performing nearest neighbor clustering on the event data according to the clustering threshold value to obtain a clustering result containing a detection result M is the total number of event points contained in the clustering result; If the clustering result comprises a plurality of categories, calculating two-dimensional space coordinate deviation of each category and the rest categories, and if the two-dimensional space coordinate deviation does not exceed the two-dimensional position coordinates in the clustering result, uniformly interpolating at a fracture position, and supplementing a fracture track to obtain a preliminary detection result; And performing point cloud filtering on the preliminary detection result to obtain a final detection result.
2. The sparse event point-oriented space-time clustering small target detection method according to claim 1, wherein the preprocessing of the original data shot by the event camera to obtain the preprocessed event data comprises the following specific implementation processes: Removing four-dimensional event point coordinates in original data Downsampling the event point coordinates after removing the polarity to obtain three-dimensional space-time point cloud data The three-dimensional space-time point cloud data is event data after preprocessing; For the two-dimensional position coordinates of the imaging pixel, For the imaging time, N is the number of event points in the original data.
3. The sparse event point oriented space-time clustering small target detection method according to claim 1, wherein the implementation process of preprocessing the original data shot by the event camera comprises the step of performing downsampling operation on the original data.
4. The sparse event point oriented space-time clustering small target detection method according to claim 1, wherein the event data are sliced to obtain a plurality of data segments, a weight value is assigned to each event point in each data segment, and the event points of all the data segments are arranged in descending order according to the weight value.
5. The sparse event point oriented spatio-temporal clustering small target detection method of claim 1 or 4, wherein any event point Weight value of (2) The calculation formula of (2) is as follows: Wherein For event points And event point The euclidean distance between the furthest event points, Representing event points And the first The Euclidean distance of each event point, m is the total number of input events.
6. The sparse event point oriented space-time clustering small target detection method according to claim 5, wherein the event points with weight values distributed below 50% are selected, and the average weights of the selected event points are calculated to obtain the adaptive clustering range 。
7. The sparse event point-oriented space-time clustering small target detection method according to claim 1, wherein the specific implementation process of performing the point cloud filtering on the preliminary detection result comprises the steps of calculating a time span of each clustering category, and deleting the clustering category if the time span of a certain clustering category does not exceed 10% of a point cloud time dimension.
8. Terminal device comprising a memory, a processor and a computer program stored on the memory, characterized in that the processor executes the computer program for carrying out the steps of the method according to one of claims 1 to 7.
9. A computer-readable storage medium, on which a computer program/instruction is stored, characterized in that the computer program/instruction, when executed by a processor, implements the steps of the method according to one of claims 1 to 7.

Description

Sparse event point-oriented space-time clustering small target detection method Technical Field The invention relates to a small target detection technology in point cloud, in particular to a sparse event point-oriented space-time clustering small target detection method. Background While this approach has met with some success, the prior art detects small targets by reconstructing the event camera data into intensity images, the method of reconstructing the intensity images loses the advantage of the event camera in terms of time resolution. In addition, some technologies perform target detection by projecting event camera data to adapt to an image-based detection method, but such technologies mainly utilize edge information of the event data instead of space-time information, and a condition that detection is missed due to lack of space-time information occurs in detection. Most techniques for object detection using event cameras employ convolutional neural network methods, which use large amounts of labeled data to train a network model to extract temporal and spatial features of event camera data. However, such techniques typically require a significant amount of preparation due to the lack of event data sets and the large computational requirements. In recent years, a method for performing target detection by using a pulse neural network to process event data has been successful in extracting time characteristics while reducing the calculated amount, but the presently disclosed sparse event point data set for the target detection field is few and insufficient in volume, and no disclosed data set for small target detection exists, so that the problem of lack of the data set is not solved in the prior art. The existing technology for detecting the targets in the event camera does not have an algorithm specific to small targets, the algorithm for sparse event point data is mostly specific to tangible targets such as pedestrians, buildings and automobiles, the detection effect on targets without the shapes is poor, and the detection precision is low. While the method for the image can detect the target position, only the horizontal position within a certain time range can be detected according to the frame frequency, and the time precision is lost. In the prior art, by detecting edge characteristics such as a plane, a curved surface and the like in the event point cloud, the methods have good effects on the object with the shape, but the object without the shape exists in the event point cloud in a straight line and a curve mode, so that the detection problem of the object without the shape is difficult to solve in the existing method. The terms used in the present invention are explained as follows: The event camera, also called neuromorphic vision sensor and bionic silicon-based vision sensor, is a device with a new imaging system, each pixel works independently, and outputs asynchronous space-time pulse signals when brightness changes, and the original data form of the event camera is as follows: Zi=(x,y,t,p); wherein Z i is the output of the ith pixel sensor, (x, y) is the two-dimensional position coordinate of imaging pixel i, t is the imaging moment (different among pixel sensors), and p is the event polarity (the light intensity increases or decreases correspondingly to positive and negative polarities). Compared with the traditional camera, the event camera has microsecond time resolution, is good at capturing a target moving at a high speed, and can overcome motion blur when the traditional camera shoots the target moving at the high speed. Furthermore, event cameras have a very high dynamic range so that they can still function properly under challenging lighting conditions. The event point cloud is obtained by removing the polarity of the original data shot by the event camera, has two-dimensional space and time dimension information, and lacks depth information compared with the traditional laser radar point cloud. The conventional point cloud is composed of points in three directions (x, y, z) at the same time, and the event point cloud is formed by the increase of the light intensity variation of the two-dimensional space (x, y) along with the time t. The small target is usually a target with a resolution of less than 32×32 pixels or a pixel ratio of less than 2% of the whole image, the small target in the event camera is not clearly defined at present, the definition in the image is adopted in the invention, the resolution of the camera photographed in the embodiment of the invention is 1920×1080, the pixels triggered by the target in 3ms are about 10-100 different, and the requirement of 2% is met. Disclosure of Invention The invention aims to solve the technical problems of large difficulty in detecting small targets in event point cloud and difficult extraction of target features, and provides a sparse event point-oriented space-time clustering small target detection method. In order to s