Researchers have developed an electronic image stabilization algorithm for shipborne cameras that enhances video stability by coupling neural feature extraction with vessel motion compensation. In maritime navigation, cameras mounted on ship platforms provide critical sensory data for crew monitoring and automated decision-making. However, dynamic marine environments subject shipborne cameras to continuous, erratic disturbances that severely degrade visual clarity. In a study published in the Journal on Image and Video Processing, Jiexuan Zhuang and colleagues introduced an end-to-end framework addressing these challenges through three synchronized stages: feature point recognition, optical flow motion estimation, and fixed-window trajectory smoothing. Experimental validation demonstrated that the proposed method reduced mean square error by 44 to 64 percent while increasing peak signal-to-noise ratio by 9 to 17 percent across diverse sea trials. By maintaining video frame consistency without incurring heavy computational costs, the system offers an effective solution for marine vision systems operating under demanding nautical conditions.
The operational environment aboard seagoing vessels presents unique physical hurdles that conventional video stabilization systems struggle to resolve. Cameras mounted on vessel superstructures are constantly disturbed by turbulent winds, ocean waves, swells, and high-frequency structural vibrations generated by onboard machinery. These disturbances produce pronounced lateral shifts, angular displacements, jitter, and motion blur across captured video sequences, obscuring navigational hazards and degrading situational awareness. Prior research explored various stabilization strategies across mobile platforms. Shang utilized low-rank constraint regularization and path prediction smoothing, achieving a 2.18-decibel improvement in image time fidelity and a 5.7 percent increase in structural similarity. Li integrated grid motion statistics with adaptive Kalman filtering to reduce blur entropy, while Wang implemented a quaternion-based self-supervised sparse optical flow transformer. In maritime settings, Gao explored motion compensation decomposition, and Xin deployed a three-degree-of-freedom gimbal combined with electronic sea-sky horizon cropping. Despite these advancements, existing shipborne cameras still suffer from multi-angle drift, environmental complexity, loss of detail, and computational latency.
To overcome the limitations of prior techniques, Zhuang and co-authors designed a modular electronic image stabilization architecture specifically engineered around the physical dynamics of marine vessels. The system operates by calculating temporal motion trajectories of salient image feature points across successive video frames, using these dynamic vectors as the computational foundation for motion compensation and frame interpolation. Rather than treating stabilization as an isolated smoothing filter, the authors structured their pipeline into three integrated components: a feature recognition module, a motion estimation module, and an image smoothing module. In the initial stage, an optimized deep learning model identifies reliable visual keypoints across the scene while discarding uninformative regions. The motion estimation stage subsequently processes these detected points through an accelerated optical flow network to derive precise camera displacement vectors over time. Finally, the image smoothing module applies a fixed-window weighting algorithm to filter mechanical jitter from genuine vessel trajectories, concluding with composite affine transformations that warp distorted frames into a stable video sequence.
The initial stage of the pipeline relies on a dedicated feature point detection architecture derived from the MagicPoint network framework. To satisfy real-time constraints required for maritime perception, the initial segment of this network utilizes a Visual Geometry Group neural network backbone characterized by extensive parallel processing capabilities. The backbone consists of sequential convolutional layers, nonlinear activations, and three max-pooling operations that progressively reduce spatial dimensions while preserving essential structural representations. A notable challenge encountered during maritime feature extraction is visual interference caused by homogeneous sky regions and foaming waves around the vessel hull. To focus network attention on informative visual landmarks, the researchers integrated an Efficient Channel Attention mechanism directly behind the convolutional blocks of the encoder. Following feature extraction, a Softmax activation function and subsequent tensor reshape operations restore the intermediate representation to original image dimensions, yielding dense feature points concentrated primarily along textured marine surfaces below the horizon line.
The technical refinement of the feature extraction network involved targeted enhancements to both the attention mechanism and the learning loss formulation. To capture cross-channel interdependencies without incurring excessive computational complexity, the team introduced an ECA module after the last two convolutional layers of the encoder. The ECA mechanism applies global average pooling to yield a 1-by-1-by-C vector, which undergoes adaptive one-dimensional convolution of kernel size k followed by Sigmoid activation. Kernel size k varies with channel dimension C according to C equals phi of k, which equals two raised to the quantity gamma times k minus b. With gamma equal to two and b equal to one, kernel size k is determined by the nearest odd integer of log base two of C divided by gamma plus b divided by gamma. Furthermore, to counter sample imbalance from disruptive pixels around the ship hull, the authors combined cross-entropy loss with focal loss across an eight-by-eight grid of feature vectors, each containing sixty-five response values.
Rigorous ablation experiments confirmed that these architectural modifications delivered measurable gains in feature point extraction precision without burdening processing capacity. Testing diverse configurations showed that appending the ECA module after the three-by-three convolutional layers in the final two encoder blocks achieved an optimal mean average precision of 93.7 percent, as deploying modules earlier expanded parameter counts and FLOPs without equal benefit. Incorporating focal loss further improved detection accuracy by 5.2 percent without altering the model footprint of 1,302,000 parameters and 6,457 million FLOPs, achieving an overall validation accuracy of 94.5 percent compared to baseline MagicPoint. In comparative benchmarks against classical feature detectors, the optimized network demonstrated notable precision. Conventional algorithms like Shi-Tomasi recorded a mean error rate of 1.933 percent with excessive spurious detections, while SIFT exhibited a 1.715 percent error rate across nearly 300,000 points. Compared to MagicPoint, the authors’ network decreased mean error from 0.987 percent to 0.856 percent, successfully excluding featureless sky regions in favor of textured marine surfaces below the horizon.
Following feature extraction, the stabilization pipeline determines inter-frame motion vectors using an optimized version of the Recurrent All-Pairs Field Transforms optical flow network. Conventional RAFT architectures compute motion vectors between consecutive frames by executing all-pairs dot products between extracted feature tensors, generating exhaustive correlation volumes. However, the authors observed that maritime video streams recorded at twenty-five frames per second exhibit small inter-frame intervals of forty milliseconds, during which spatial displacement between adjacent frames remains constrained. Performing exhaustive all-pairs matrix multiplications under such conditions introduces severe computational redundancy that degrades real-time efficiency. To resolve this inefficiency, the researchers incorporated a k-nearest neighbor algorithm into the motion estimation framework. Rather than matching each feature vector against the entirety of the subsequent frame, the k-NN mechanism identifies only the k closest feature vectors between consecutive frames. This targeted matching strategy restricts the resulting similarity matrix to dimensions of height times width times k, significantly decreasing processing latency and computational overhead while maintaining temporal continuity.
To validate the motion estimation component, the authors conducted comparative evaluations benchmarking their improved RAFT model against baseline RAFT, PWC-Net, FlowNet1, and FlowNet2 across inference latency, parameter count, and endpoint error. The endpoint error was quantitatively calculated as the Euclidean distance between estimated optical flow coordinates and ground-truth motion vectors. Experimental results indicated that FlowNet1 and the improved RAFT model achieved the shortest inference times among all evaluated architectures, whereas PWC-Net exhibited the longest processing latency. In terms of architectural compactness, the improved RAFT possessed the smallest parameter footprint, requiring approximately fifty times fewer parameters than FlowNet2. The authors concluded that the improved RAFT architecture struck the optimal balance between inference speed, computational economy, and tracking fidelity, making it the most suitable engine for shipborne motion vector estimation.
The third component of the framework processes extracted displacement vectors to generate stabilized video sequences through fixed-window weighted trajectory smoothing and affine motion compensation. Graphical analysis of vessel motion revealed that rotational fluctuations remained minimal over time, forming a near-horizontal trajectory that allowed smoothing calculations to focus primarily on vertical translation vectors. To distinguish erratic camera jitter from intentional vessel maneuvers, the algorithm evaluates the absolute first derivative between consecutive frame coordinates, denoted as Diff. Trajectory regions with small derivatives represent smooth, stable motion, whereas large derivatives reflect pronounced jitter. The algorithm derives an adaptive weight W by multiplying the reciprocal of Diff by a correlation coefficient q, subsequently computing the smoothed vector coordinate as a weighted combination of the prior and current frames. The system then builds a composite affine transformation matrix incorporating translation amounts dx and dy with rotation angle theta, mapping displaced frames back into stable alignment. Trajectory comparisons demonstrated that fixed-window weighting delivered higher accuracy and smoother paths than standard Kalman filtering.
The complete electronic image stabilization framework was subjected to comprehensive empirical validation using authentic maritime video sequences captured near the coast of Xiamen City. The experimental dataset comprised video footage recorded from the bow viewpoint of an operational vessel at a resolution of 848 by 480 pixels and a frame rate of 25 frames per second. To assess performance across challenging operational scenarios, the researchers selected ten representative video segments of 200 consecutive frames each, encompassing heavy sea motion, abrupt navigational maneuvers, and low-texture maritime scenes. Quantitative evaluation demonstrated that the proposed algorithm decreased mean square error by approximately 44 to 64 percent while increasing peak signal-to-noise ratio by 9 to 17 percent compared to unstabilized video sequences. By uniting lightweight neural feature detection, accelerated optical flow estimation, and adaptive trajectory smoothing, this research provides a robust foundation for reliable video perception in intelligent maritime navigation systems.
Subject of Research: Electronic image stabilization for shipborne cameras based on feature extraction and vessel motion
Article Title: An electronic image stabilization algorithm for shipborne cameras based on image feature points and ship movement
Article References: Zhuang, J., Wang, C., Chen, Z., Li, Y., Wu, R., Zhu, S., Guan, L., Liu, L., & Man, W. (2026). An electronic image stabilization algorithm for shipborne cameras based on image feature points and ship movement. Journal on Image and Video Processing, 2026(1), Article 8. https://doi.org/10.1186/s13640-026-00694-7
Image Credits: AI Generated
DOI: 10.1186/s13640-026-00694-7
Keywords: electronic image stabilization, shipborne cameras, maritime navigation, feature point detection, motion estimation, optical flow, RAFT model, k-nearest neighbor, efficient channel attention, MagicPoint, trajectory smoothing, affine transformation, mean square error, peak signal-to-noise ratio, focal loss, vessel motion, visual geometry group
Cite Scienmag News
Blake Davidson. (November 18, 2026). Shipborne Camera Algorithm Stabilizes Maritime Video Through Feature Tracking and Trajectory Smoothing. Scienmag. https://scienmag.com/shipborne-camera-algorithm-stabilizes-maritime-video-through-feature-tracking-and-trajectory-smoothing/
Blake Davidson. "Shipborne Camera Algorithm Stabilizes Maritime Video Through Feature Tracking and Trajectory Smoothing." Scienmag, 18 November 2026, https://scienmag.com/shipborne-camera-algorithm-stabilizes-maritime-video-through-feature-tracking-and-trajectory-smoothing/. Accessed 18 November 2026.
Blake Davidson. "Shipborne Camera Algorithm Stabilizes Maritime Video Through Feature Tracking and Trajectory Smoothing." Scienmag. November 18, 2026. https://scienmag.com/shipborne-camera-algorithm-stabilizes-maritime-video-through-feature-tracking-and-trajectory-smoothing/








