The results come from three 60-second football video sequences and do not establish the system as a replacement for match officials. The study focused on automatic identification and tracking of players and the ball in soccer video. Its offside framework compares attacking-player positions with the ball and the offside line associated with the second-placed defending player.

The researchers followed the 2023 FIFA rule described in the paper, under which the determination considers the player’s head, torso and feet but not the arms. The system was tested in three deliberately contrasting scenes: a rapid attack-to-defense transition in daylight, an evening sequence with gradually dimming light and shadows, and a brightly illuminated night counterattack with high background contrast and slight camera shake. At the center of the detection method is a five-component Gaussian mixture model, a background-subtraction technique that represents the changing appearance of each pixel with multiple Gaussian distributions.

The researchers set the number of components to five to balance the difficulty of modeling complex football backgrounds against computational cost. Conventional GMM systems can generate false detections, missed foreground regions and broken edges when illumination changes or moving background elements interfere. A fixed learning rate creates a second problem: if updates are too slow, the model cannot adapt to environmental changes and may leave ghost traces around fast-moving players; if updates are too rapid, slowly moving objects can be incorporated into the background.

The proposed GMM adjusts its learning rate using both the stage of model construction and measured changes in the scene. During the first 100 frames, the initial rate of 0.10 is increased gradually to accelerate convergence and suppress early interference. In later frames, the base rate is 0.05 and the minimum permitted value is 0.02.

The system estimates illumination change from the mean normalized grayscale difference between the current frame and the background model. When this measure exceeds 0.15, the background update is accelerated; when illumination stabilizes, the rate returns toward its base value. In a sensitivity analysis, the improved model at a rate of 0.06 produced 90% accuracy, 87% recall and an 89% F1 score, while the corresponding fixed-rate GMM began to deteriorate as the rate increased.

To recover motion regions more completely, the researchers used four-frame differencing rather than conventional three-frame differencing. Four consecutive frames yield three binary difference images. The system combines their union, representing all detected inter-frame changes, with their intersection, representing changes visible across all three comparisons.

Each difference image is binarized using an Otsu-derived threshold plus a correction based on the image’s grayscale standard deviation; the light-suppression coefficient for that correction is 0.5. When the analysis window increased from three to four frames, average detection accuracy rose by 4.7% in the daylight scene, 7.3% in the dimming evening scene and 9.3% in the night counterattack. False-detection rates fell by 3.9%, 4.3% and 4.9%, respectively.

Adding fifth and sixth frames produced less than 0.5% additional accuracy in each scene, while the false-detection rate tended to saturate and per-frame processing time increased. The four-frame configuration was therefore selected as the balance between detection performance and computational efficiency. The paper also fuses Canny edge information with the motion mask to address holes and incomplete outlines.

The upper Canny threshold is selected adaptively through Otsu’s method, and edges near the foreground are isolated by intersecting them with a foreground mask dilated using a 3-by-3 kernel. In the ablation study, four-frame differencing contributed about 7.1% to average accuracy relative to three-frame differencing, while Canny fusion added about 1.0%. The full detection configuration reached 95.4% average accuracy, compared with 85.9% for traditional GMM with three-frame differencing.

Detection supplies positions in individual frames, so the study added tracking to preserve each target’s identity through time. The baseline was Fast Discriminant Scale Space Tracking, or FDSST, which learns a multichannel correlation filter and locates a target in later frames by finding the strongest response. Principal component analysis reduces the dimensionality of the features used by the filter, while incremental updates allow the model to adapt to the video sequence.

The researchers selected FDSST because it can adjust to changes in target scale without requiring a separate deep re-identification network, additional training data or large annotated datasets. In comparisons over frame positions from 5 to 225, the improved GMM-FFD detector ranged from 85.4% to 99.3% accuracy, versus maximum accuracies of 83.7% for ViBe, 82.9% for GMM and 59.7% for kernel density estimation. Severe occlusion remains difficult for correlation-filter tracking because the response map may develop several peaks or its strongest peak may fall sharply, causing drift or loss.

The modified FDSST therefore introduces a Kalman-filter fallback governed by a confidence score derived from the maximum, mean and standard deviation of the response map. Above 0.65, the FDSST estimate is treated as reliable; below 0.35, the system switches to the Kalman filter and predicts the target from its horizontal and vertical positions and velocities. Within that interval, it retains the mode used in the previous frame.

In the occlusion test, mean center-point error was 20.2 pixels for the dynamically switched tracker, compared with 29.3 pixels for baseline FDSST and 26.0 pixels when the Kalman filter was always active. The improved tracker’s accuracy curve stabilized at 86.7%, versus 73.4% for the unmodified version. The experimental protocol separated parameter tuning from testing.

The three ISSIA-CNR sequences, each 60 seconds long, contained about 4,500 frames at 1920 by 1080 pixels and 25 frames per second. Two experts manually annotated player bounding boxes and offside lines, and their agreement was evaluated with a Kappa coefficient of 0.92. For each sequence, the first 70% of the frames were used for grid-search parameter tuning and the final 30% for testing.

Evaluation covered average accuracy, false-positive rate and frames per second, with results averaged across three runs. Tracking error was sampled every 40 frames. SoccerNet was used only for cross-dataset validation, while comparison methods followed the same experimental protocol.

Against deep-learning detectors under the reported setup, the hybrid detector processed 35 frames per second on ISSIA-CNR, compared with 23 frames per second for YOLOv5s and 12 frames per second for Faster R-CNN. Its mAP@0.5 of 92.9% was below the reported 93.1% for Faster R-CNN but above 91.6% for YOLOv5s and 92.4% for YOLOv8s. The complete tracking system ran at 35 frames per second with a reported 68.4% Multiple Object Tracking Accuracy and 28 identity switches.

The method has 59.41 million parameters and still requires substantial computing resources. Heavy rain is another stated limitation: rain streaks and ground reflections can interfere with target-feature extraction and increase detection errors. The authors identify region-of-interest narrowing, feature-dimension reduction, less frequent detection during low-risk periods and a lightweight rain-suppression module as areas for further work, but those steps were not evaluated in the reported experiments.

Subject of Research: Computer-vision detection and tracking of soccer offside positions in match video

Article Title: Automatic offside determination in soccer videos based on improved GMM-FFD-FDSST

Article References: Zhang, C., Dai, J., & Liu, Y. (2026). Automatic offside determination in soccer videos based on improved GMM-FFD-FDSST. Discover Artificial Intelligence, 6(1), Article 1199. https://doi.org/10.1007/s44163-026-02046-w

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02046-w

Keywords: soccer offside, computer vision, video analysis, Gaussian mixture model, four-frame differencing, Canny edge detection, FDSST tracking, Kalman filtering, occlusion handling, sports analytics, Automatic, offside

The work, published in the journal Applied Intelligence, is built around two guiding ideas that give the method its name: dynamic adaptation and collaborative enhancement. The authors argue that current tracking systems tend to break down in three recurring situations. First, objects change pose dramatically as they move, so the appearance a tracker memorized a second ago no longer matches what the camera sees now. Second, in dense crowds, many objects look nearly identical, which confuses the appearance-based matching that most modern trackers rely on. Third, motion patterns are rarely the smooth, linear trajectories that classical tracking mathematics assumes; people stop, sprint, pivot, and jostle, and vehicles brake and swerve. When these difficulties stack up in a crowded scene with frequent occlusions, trackers suffer from identity switches, where one person’s label is handed off to another, and from track loss, where an object simply disappears from the system’s bookkeeping.

The first pillar of the new method is a nonlinear adaptive Kalman filter. The Kalman filter, a mathematical tool dating back to the Apollo era, is the workhorse of motion prediction in tracking. It maintains a running estimate of an object’s state, typically its position and velocity, and predicts where the object should appear in the next frame, then corrects that prediction with the latest detection. The problem is that the standard formulation assumes motion is well behaved. When a pedestrian suddenly stops or reverses direction, the filter’s prediction drifts far from reality, and the association step that matches detections to tracks starts making mistakes. The new approach equips the filter with a nonlinear adjustment mechanism that detects anomalous motions, situations where the observed position deviates sharply from what the linear model expects, and adapts the state update accordingly. In effect, the filter learns when its own assumptions are failing and recalibrates the prediction step to better capture the object’s true dynamics rather than forcing every object through the same rigid linear model.

This matters because motion prediction and appearance matching are deeply coupled in a tracking pipeline. Most trackers decide which detection belongs to which track by combining a motion cost, how far the detection is from the predicted location, with an appearance cost, how similar the visual features are. If the motion prediction is wrong, the tracker may reject the correct detection as too distant and instead grab a nearby wrong one, producing an identity switch. By making the prediction step nonlinear and adaptive, the framework reduces these cascading errors at their source. The authors describe this as optimizing the prediction step so that it captures object dynamics rather than merely extrapolating past positions, a change that proves especially valuable in scenes where people move erratically or where the camera itself introduces complex relative motion.

The second pillar addresses the appearance side of the problem with what the team calls a multi-dimensional feature enhancement network. Appearance features are typically extracted by a deep neural network that converts each detection into a compact descriptor, sometimes called a re-identification or re-ID embedding, which should remain stable for the same person across frames. In practice, these descriptors are fragile: low-resolution video, motion blur, poor lighting, and compression artifacts all degrade the input image, and the resulting features become unreliable, especially for small or partially visible objects. The enhancement network attacks this by accumulating appearance information across multiple scales, blending fine-grained detail with coarser contextual cues so that the descriptor for an object is less dependent on the quality of any single crop of the input image. The authors report that this cross-scale appearance-enhancing exploration reduces the influence of input image quality, which is precisely the kind of robustness needed for surveillance footage and dashcam video, where resolution and lighting are rarely ideal.

Even with better motion prediction and better appearance features, occlusion remains the great destroyer of tracks. When two people cross paths in a crowded plaza, one is briefly hidden behind the other, and the tracker must decide whether the detection that reappears on the far side belongs to the track it lost or to a brand-new object. To handle this, the framework introduces a third component: a trajectory stitching network. Rather than treating lost tracks as dead, the system keeps the fragments and compares their spatiotemporal similarity, examining both where and when the fragments begin and end. If a fragment that disappeared near a doorway reappears moments later at a plausible location with a plausible time gap and a matching appearance, the stitching network merges the pieces into one continuous track. This is a direct assault on the identity switches that dominate error statistics in crowded-scene benchmarks, and it reflects a broader trend in the field, visible in methods like OC-SORT and StrongSORT, of treating occlusion handling as a first-class design goal rather than an afterthought.

The researchers evaluated their method on MOT17 and MOT20, the canonical public benchmarks maintained by the MOTChallenge community. MOT17 contains pedestrian videos captured from both static and moving cameras in a range of lighting conditions, while MOT20 pushes the difficulty further with extremely dense crowds, in some frames containing well over a hundred simultaneous people. These datasets are deliberately unforgiving: they include the exact combination of occlusion, similar appearance, and irregular motion that breaks weaker trackers. According to the paper, the proposed approach outperforms other advanced tracking approaches on both benchmarks, meaning it achieves better scores on the standard metrics that trade off tracking accuracy, identity preservation, and false positives. The authors also note that the datasets used in the study are publicly available through the MOTChallenge repository, and they state that they plan to release the code after acceptance, which would allow other groups to build on and verify the results.

The significance of this work lies less in any single component than in the way the components collaborate, which is what the authors mean by collaborative enhancement. A better Kalman filter alone cannot fix bad appearance features, and better features alone cannot rescue a track that has been lost for twenty frames. By jointly improving motion modeling, appearance representation, and fragment reconnection, the framework attacks the identity-switch problem from three directions at once. This systems-level view is increasingly common in the tracking literature, where recent entries such as ByteTrack, MotionTrack, BoostTrack, and various transformer-based trackers have each pushed different levers, from associating every detection box regardless of confidence to learning long-term motion patterns. The new method’s contribution is a coherent architecture in which dynamic adaptation keeps the motion model honest while the enhancement networks keep the visual evidence strong enough to stitch through occlusions.

The applications extend well beyond the benchmark videos. The authors point to intelligent transportation, security, and augmented reality as the domains where multiobject tracking is crucial. In traffic monitoring, robust tracking underlies everything from counting vehicles to predicting collisions; in security, it enables behavior analysis across crowded stations and stadiums; in augmented reality, virtual objects must remain anchored to real people and vehicles as they move and occlude one another. Aerial and drone-based tracking, an area explored by related systems such as RAMOTS, and sports analytics, where methods like Deep HM-SORT have targeted occlusion-heavy game footage, would also stand to benefit from trackers that survive dense, chaotic scenes. The work was supported in part by China’s National Key Research and Development Program and by Shandong Province research grants, reflecting the substantial national investment in computer vision infrastructure.

There are, of course, the usual caveats. The published results are benchmark numbers, and real-world deployment brings domain shifts, edge cases, and computational constraints that leaderboards do not capture. The authors state that code release is planned, so independent replication will be the next test. Still, the trajectory of the field is clear: trackers are moving from rigid linear assumptions toward adaptive, nonlinear motion models, and from single-cue matching toward multi-dimensional, occlusion-aware architectures that treat a lost track as a puzzle to be solved rather than a failure to be logged. If the gains reported on MOT17 and MOT20 hold up outside the lab, the crowded, blurry, occlusion-riddled videos that once defeated tracking systems may finally become tractable, and the machines watching the world’s busiest places may stop losing track of the very people they are meant to follow.

Subject of Research: A multiobject tracking method using dynamic adaptation and collaborative enhancement for crowded scenes

Article Title: A multiobject tracking method based on dynamic adaptation and collaborative enhancement

Article References: Ji, Y., Guo, Y., Liu, Z., Li, H., Liu, Z., & Fang, H. (2026). A multiobject tracking method based on dynamic adaptation and collaborative enhancement. Applied Intelligence, 56(14), Article 412. https://doi.org/10.1007/s10489-026-07438-0

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07438-0

Keywords: multiobject tracking, computer vision, Kalman filter, occlusion handling, trajectory stitching, appearance features, MOT17, MOT20, deep learning, motion prediction, identity switches, surveillance