Automatic extraction channel of space debris based on wide-field surveillance system

Space activities since 1957 have generated substantial debris that threatens spacecraft and satellite operations, necessitating continuous monitoring for collision avoidance and orbit prediction. Detecting and tracking dim space targets in wide-field optical images is challenging due to weak target intensity, small apparent size, dense star backgrounds, and noise. Prior approaches include template matching, morphological operations, neural networks, and 3D matched filtering; however, they often suffer from sensitivity to unknown target speeds, background clutter, or high computational cost for multiple hypothesis tracking in dense scenes. This study focuses on improving automated detection and tracking of streak-like debris targets in complex, star-rich backgrounds using a wide-field CMOS telescope in star-tracking mode.

The paper surveys several lines of work for dim/small target detection in cluttered imagery: (1) Template/matched filtering methods, including 3D matched filtering, which perform well when target speed is known but degrade with unknown or varying speeds (Reed et al.). (2) Mathematical morphology-based methods that suppress noise/background fluctuations and can enhance detectability but may be sensitive to structural element choices (Bai et al., Sun et al.). (3) Multistage hypothesis testing (MHT) frameworks for multi-target detection using tree-structured trajectories, which can reduce missed alarms but incur exponential growth in branches as candidate starts increase, leading to heavy computation (Blostein & Huang). (4) Additional related works in hyperspectral detection, centroiding enhancement, infrared point-target detection via improved template matching, and Bayesian methods are referenced as context. The limitations highlighted include performance drops under unknown motion parameters, susceptibility to background/star density and noise, and computational burdens in wide-field surveillance with many potential targets.

Data and imaging setup: Astronomical image sequences (4096×4096 pixels) were acquired by a wide-field CMOS telescope (aperture 280 mm, focal length 324 mm, pixel size 9×9 µm, field of view 6.5°×6.5°, read noise 3.7 e-, exposure 2 s, frame rate 0.5 Hz). The system operates on a star-tracking mount, so stars appear as point sources while moving space objects form streaks. Model: Each frame f(x,y,n) is modeled as the sum of target O, stars S, background B, and noise N. Pipeline overview (Fig. 3): Three stages—(1) image preprocessing and background suppression; (2) debris detection; (3) debris tracking. Stage 1: Image preprocessing and background suppression. - Improved median filtering: Apply a 5×5 median filter to obtain a roughly denoised image. Compute its average and subtract from the original to derive a threshold identifying impulse-noise contaminated pixels. Only those pixels are replaced by the 5×5 median-filter outputs, leaving uncontaminated pixels unchanged, thereby reducing dependence on window size. - Background estimation: Use the median of five consecutive frames to estimate the sky/background ADU, then subtract this background from the median-filtered image to yield an image containing debris, stars, and residual noise. - Guided filtering: Apply guided filter with the background-suppressed image g both as input and guide to preserve edge information of streak targets while smoothing stars and isolated noise. The local linear model q_i = a_k g_i + b_k within window W_k is optimized with regularization ε to prevent overfitting. Parameter selection is critical: with fixed ε, window radius h affects results. Empirically, h = 0.3H (H = image height) and ε = 0.04 provided good suppression; too small h leaves residual stars/noise; too large h can remove debris. Stage 2: Debris detection. - Binarization via maximum between-class variance (Otsu) to reduce data and improve Hough precision. - Hough transform in polar coordinates maps collinear points (streaks) to peaks/intersections in parameter space. For a line through (x,y), the relation is p = x cos θ + y sin θ; peaks identify candidate streaks, from which endpoints and centers of line segments are derived. Stage 3: Debris tracking. - Kalman filter tracking assumes approximately uniform linear motion between adjacent frames in the wide-field optical system. Initialization uses Hough detections (initial state/position) and the first two frames for velocity. Standard prediction and correction steps are used with state transition A, control B, process noise Q, observation matrix H, observation noise R, and covariance P, yielding real-time updates of target state X_n and covariance P_n. Pseudocode (Table 2) outlines initialization and iterative update using Eqs. (7–8). Experiments: Conducted on real star-tracking sequences; visualizations show detection of two targets (T1, T2) across consecutive frames, Kalman tracking overlays and trajectories without target loss. Robustness is shown on four types of backgrounds (dense stars/noise, lower noise, cloud interference). Comparative evaluation with MHT, NTH, and IMVP uses published standard parameters.

• The proposed pipeline achieves high detection accuracy and efficiency on real astronomical images from a wide-field, star-tracking telescope. - Quantitative comparison (Table 3) on the same images: OURS detection probability 96.2%, false alarm rate 8.4%, running time 4.32 s; MHT 87.7%, 17.2%, 36.94 s; NTH 90.6%, 69.1%, 6.8 s; IMVP 91.4%, 15.9%, 10.54 s. - The method successfully detects and continuously tracks targets (e.g., T1 and T2) over multiple consecutive frames without loss, with centroid trajectory estimates closely matching true trajectories (Fig. 9), and maintains performance across varying backgrounds (including dense stars, noise, and cloud interference). - Parameter study shows guided filter window size h critically affects outcomes: h = 0.3H balances star/noise suppression and target preservation in tested data.

The study addresses the challenge of detecting dim, streak-like space debris in star-rich, noisy wide-field optical images. By combining multi-frame median background suppression with guided filtering, the pipeline effectively removes uneven sky background, stars, and isolated noise while preserving streak edges. Hough transform robustly localizes linear debris streaks, and Kalman filtering provides real-time, low-computation tracking assuming near-uniform motion between frames. The approach does not rely on auxiliary space-target catalogs or prior orbital data and integrates detection with astronomical positioning. Comparative experiments demonstrate improved detection probability, reduced false alarms, and faster runtime relative to MHT, NTH, and IMVP, indicating practical utility for distributed wide-field surveillance systems.

The paper presents an automatic extraction pipeline for space debris in wide-field, star-tracking optical imagery that integrates improved median filtering, multi-frame median background estimation, guided filtering, Hough-based streak detection, and Kalman tracking. Experiments on real data show accurate detection and robust, real-time tracking in complex backgrounds with superior accuracy-speed trade-offs compared to established methods.

The method’s performance depends on parameter choices in guided filtering: with a fixed regularization, the filter window size h strongly influences results—too small leaves residual noise and bright stars; too large can remove debris targets. The study reports optimal performance at h = 0.3H on tested images. The evaluation is conducted on star-tracking mode data from a specific wide-field telescope; other acquisition modes are not assessed in this work.