Key Technologies of Star Sensors in Complex Environments

Star map denoising algorithm of star sensors:

The discussion is about denoising methods for single pixel noise. The traditional engineering processing methods mainly use spatial filtering denoising algorithms. The main mechanism of spatial filtering denoising algorithms is smooth filtering, which includes mean filtering, box filtering, median filtering, and Gaussian filtering. These types of filtering algorithms usually have a smoothing effect on the background.

(1) Mean filtering

Mean filtering, as the name suggests, is to obtain the pixel mean of an image area, taking the current pixel as the center point, and calculate the surrounding 𝑛 ×  The statistical average of each pixel within the 𝑛 box. The advantage of mean filtering is that its calculation speed is particularly fast, but the resulting image is relatively blurry. The larger the filtering kernel, the more blurry the image becomes. Mean filtering has a good processing effect on Gaussian noise, but its effect on salt and pepper noise is average.

(2) Box filtering

The advantage of mean filtering is that its calculation speed is particularly fast, but the resulting image is relatively blurry. The larger the filtering kernel, the more blurry the image becomes. Mean filtering has a good processing effect on Gaussian noise, but its effect on salt and pepper noise is average.

(3) Median filtering

Median filtering is the process of obtaining the median value of other pixels in the neighborhood of the current central pixel to replace the current central pixel. Median filtering has a relatively good filtering effect on salt and pepper noise.

(4) Gaussian filtering

The principle of Gaussian filtering is to use the filtering kernel function for convolution operations.

The above-mentioned spatial filtering methods have a significant impact on the target during the denoising process, and may have fuzzy effects on the target during the smooth filtering process, thus having some drawbacks.

(5) Multi frame prediction method

A method of background prediction that accumulates background values through multiple frames and then removes the influence of noise through background subtraction. However, this method has requirements for the dynamic conditions of star sensors and is only suitable for low angular rates, and is mainly suitable for situations where the noise position is relatively fixed.

Scaling the correction domain at high speeds can adapt to higher dynamic speeds. As shown in Figure 1.4, during the process of star point movement, the processing frames of each frame are continuously used as background information. When the frames of multiple consecutive frames are combined into a set of backgrounds, noise removal can be achieved by background subtraction in the current frame. However, this method can still only handle fixed single pixel noise, and its suppression effect is poor for position random noise.

The algorithms mentioned above have limitations in practical applications, making it difficult to maintain the accuracy of star point centroid positioning while denoising.

Figure 1.4 The multi-frame prediction method adopted by Zheng

Star point detection algorithm

Before conducting star point recognition, it is usually necessary to obtain star point position information in advance, which is called star point detection. According to the characteristics of star maps, the main tasks of star point detection include:

(1) Segmentation between star targets and background;

(2) The segmentation between a single star target and other stars.

Star point detection is a processing method that achieves threshold segmentation of targets and backgrounds in star maps. The specific steps of star detection include clustering, segmentation, labeling, and final centroid positioning of star regions. Star detection is a crucial step before star recognition, and its accuracy plays a crucial role in the successful recognition of subsequent stars. It also determines the accuracy of recognition and the effectiveness of attitude. Therefore, a fast and effective star detection method is crucial. The clustering and segmentation of star points are generally carried out using two methods, one is the global threshold method, and the other is the local threshold method, both of which can be collectively referred to as threshold segmentation methods. When using a star sensor to aim at the sky background for shooting, it can be observed that there is a significant difference in the grayscale of the stars and the background. In this case, empirical thresholds can be used to effectively separate the stars and the background.

However, when encountering the influence of stray light, the background in the star map is very bright due to the presence of stray light, even approaching the brightness values of some weak stars, or directly flooding the star points. Therefore, using a single global threshold segmentation method is not very suitable, and it is easy to overlook one aspect and miss the other, as it cannot ensure that the star points in each region are accurately detected. Therefore, it is necessary to use a local threshold segmentation method. Among them, the commonly used methods include background prediction algorithms. However, the local adaptive threshold method is limited by the complexity of computation, often resulting in slow computation speed and difficult implementation. Moreover, a simple local threshold method cannot be applicable to various working conditions and backgrounds.

At present, most detection algorithms applied on star sensor platforms do not have ideal suppression effects on stray light. Considering that the proportion of star points on the detector target surface is usually very small, about 2 × 2 to 7 × With a size of 7 pixels, star points can also be considered as small targets, and the idea of infrared weak target detection can be used to achieve star point detection. Small object detection is usually divided into two types: based on single frame images and based on video sequences. The first is a detection method based on filtering. It is also a common method used for small object detection in the early days. Filter based detection methods have the characteristic of simple operation and achieve background suppression by designing specific filters.

In addition, combining the idea of frequency domain operation, relevant filters are designed in the frequency domain to suppress noise by utilizing the difference in frequency functions between the target and other backgrounds and noise. Overall, although algorithms have low computational complexity, their detection performance is usually average and can only be applied to background suppression in certain scenarios, with low robustness.

Filter based detection algorithms are mainly divided into the following two types:

(1) Algorithm based on spatial filtering

Early infrared dim small target detection mostly used spatial filtering algorithms, which designed filtering windows to calculate the differences between the target and background, combined with the differences in grayscale between the target and background to achieve small target detection, and also achieved background suppression; The earliest use was a simple spatial high pass filter, where the grayscale value of weak and small targets is generally higher than the grayscale value of the background. This difference is then used to achieve the final target detection. However, in this method, those isolated noise points are also easily detected, resulting in low detection rate. Later, other scholars proposed new spatial filtering algorithms, defined new open operations, and obtained a new Top-Hat transform, which can solve the problem of traditional Top-Hat transform being unable to distinguish real targets when used for infrared weak target detection.

(2) Algorithm based on frequency domain filtering

In addition to spatial filtering methods, effective filters can also be developed in the spectral transform domain to separate the target from background clutter by utilizing the frequency differences between the target, background, and clutter. Common filters for frequency domain filtering include ideal high pass filters, Gaussian high pass filters, and Butterworth filters, all of which exhibit the phenomenon of “ringing” and poor filtering performance. Later, researchers carried out some improvement work on frequency domain filtering, such as using wavelet transform methods and shear wave transform. In addition, there are methods to achieve target detection by changing the corresponding coefficients through singular values [98]. However, due to parameter setting requirements, it cannot adapt to various different scenarios. Overall, compared to algorithms based on spatial filtering, algorithms based on frequency domain filtering have higher complexity, but also have higher detection levels.

The Human Visual System (HVS) mechanism is widely used in object detection, object recognition, and behavior understanding. This type of algorithm utilizes the visual perception characteristics of the human eye and adopts local difference and mutation methods to achieve small object detection tasks. The small target detection methods based on HVS can be specifically divided into the following two types:

(1) Algorithm based on spectral residuals

Small target detection is achieved by retaining the normative characteristics of the target while suppressing other characteristics. This algorithm has good detection performance for infrared dim small targets that do not require prior knowledge and do not have significant texture features, with low algorithm complexity, but poor ability to suppress background clutter. Some experts have gradually improved the spectral residual algorithm, such as improving the threshold of variance and mean in the amplitude spectrum. The technology of combining corner detection and local feature fusion can effectively achieve object detection and improve the accuracy of detection.

(2) Algorithm based on local contrast

Considering that there is a gradient difference in the grayscale values between neighboring images and weak targets in infrared weak target images, which has significant characteristics and conforms to the characteristics observed by the human visual system, selecting local features can better represent small targets and facilitate detection. However, small object detection algorithms have some limitations in practical use and do not have complete robustness to different complex scenarios.

In summary, filter based methods are mainly suitable for single, uniform continuous backgrounds and scenes with smaller target sizes; The HVS based detection method is mainly suitable for scenes with high target brightness and significant differences from the surrounding background. Similar to star sensor star point detection methods, these algorithms have varying adaptability to targets and environments, and cannot perfectly cope with various types of stray light encountered by star sensors.

Star point recognition algorithm

Star point recognition algorithms can generally be divided into three categories:

(1) Star point recognition algorithm based on constellation features;

Firstly, star recognition algorithms based on constellation features utilize feature quantities for recognition, which are determined by the positional relationship between stars.

The earliest star recognition algorithm based on constellation features was the triangle algorithm proposed in the 1970s. As the name suggests, it utilizes the triangle features composed of stars for star recognition. Although it has been developed for many years, it and its improved methods are still the most commonly used star recognition algorithms in engineering.

Compared to the earliest methods of simply using corner distance features, the triangle algorithm has higher feature dimensions and higher recognition accuracy. The disadvantage of triangle recognition algorithm is also quite obvious, that is, when the number of matching triangles is large, it is prone to incorrect matching, which leads to a decrease in recognition success rate. In addition to the triangle algorithm, there is also a matching group based algorithm that utilizes multiple bright stars as the main star, calculates the angular distance relationship between the main star and the auxiliary star, and achieves matching search with the star library. However, due to the uncertainty of the size of the matching group, only angular distance can be used as the basic parameter, which can easily lead to uncertainty in the recognition process. The matching group algorithm is also a commonly used recognition algorithm in engineering model tasks.

Pyramid recognition algorithm based on k-vector technology: k-vector technology can achieve fast query of star catalogs, thereby reducing the search times of navigation star catalogs and ultimately improving the speed of star point recognition. The pyramid algorithm selects four observation stars to form a pyramid feature graph, with one star as the vertex, and then the remaining three stars form a triangle and form a pyramid shaped tetrahedron with it. The disadvantage is that as the star pair information table increases, it will cause a decrease in the accuracy corresponding to the fitting curve, resulting in the best matching star pair not being within the range of angular error.

(2) Star point recognition algorithm based on character pattern;

The grid recognition algorithm maps star coordinates onto a sparse matrix and achieves recognition through matrix matching. The advantages of grid algorithm are very obvious, as it has the advantages of small storage space requirements, fast recognition speed, and insensitivity to measurement errors. However, when the positioning error or corresponding magnitude error of a star is particularly large, its corresponding recognition success rate will quickly decrease. The “Elastic Gray Grid Algorithm” adds a virtual elastic template during recognition. The elastic grayscale grid algorithm can effectively improve the fault tolerance for magnitude and noise errors, thereby further improving the recognition rate. Using a circular grid instead of the initial square grid can solve the dependence of the square grid algorithm on reference stars, while also having high robustness to rotating environments. The KMP (Knuth Morris Pratt) star recognition algorithm is another manifestation of the grid algorithm. The KMP star recognition algorithm performs high pass filtering on the star map, generates 0-1 strings and corresponding formal matching patterns, and finally uses the KMP string search algorithm to complete the string recognition of the star map.

(3) A star recognition algorithm based on intelligent behavior.

This is commonly known as a recognition method based on neural networks. Scholars have introduced neural networks into star recognition and proposed a star recognition method based on fuzzy neural networks. This algorithm uses the angular distance information of the three corresponding sides that make up the triangle, then trains and learns the navigation triangle library, and finally uses the learned neural network structure to complete star map recognition. The advantages of this algorithm lie in its high recognition success rate and fast recognition speed, but its disadvantages are also obvious: slow training and learning speed, requiring a large number of samples for training, and the accuracy of recognition can easily vary due to differences in training set size and training time. Additionally, it is a significant test for the algorithm implementation platform. The star recognition method based on genetic algorithm first selects a main star, and then uses encoding to complete the classification of the star diagonal distance between the main star and its companion star, as well as the angle between stars. In addition, two sets of star maps corresponding to the star angular distance error and angle error sum are specified as fitness functions, and then coarse and fine positioning searches are carried out twice to increase the mutation factor in the coarse positioning stage, Reduce the variation factor during the fine localization stage.

The three major types of methods mentioned above are common star recognition algorithms, but most algorithms accumulate errors under the interference of transient effects and cannot achieve effective and continuous star recognition. In recent years, some experts have proposed new recognition methods for star point recognition under noise interference, but these algorithms still cannot achieve high recognition success rates under a large amount of transient noise interference.

Focusing on the key technologies of star sensors in complex environments, including star map denoising, star point detection, and star point recognition Conduct research from multiple aspects, introduce edge protection ideas, and propose a filtering sub template suitable for star images based on the characteristics of star points; Introducing the concept of curvature to characterize star maps; Based on the characteristics of transient noise, preliminary suppression of transient noise is achieved by using noise suppression methods.