The problem the team tackled is subtle but fundamental to multi-camera perception. Visual tracking—estimating the position and trajectory of a target object across consecutive video frames—underpins applications from autonomous driving to video surveillance to robotics. Ground cameras sit close to their subjects, capturing rich appearance detail but lacking flexibility; UAV cameras hover at altitudes of 10 to 50 meters in the benchmark dataset, offering a sweeping field of view and global motion context but only coarse appearance information. Fusing these complementary views should, in principle, make tracking far more robust than either camera alone. The first attempt at this fusion, a multi-view collaborative learning module known as MvCL built on the TransT transformer tracker, associated and fused features from the two views using a cross-attention mechanism driven purely by appearance similarity. Yet in practice it faltered whenever camera motion, occlusion, motion blur, or visually similar distractors entered the scene, particularly in the UAV view, where the target appears small and the wider field of view heightens sensitivity to disturbance.

The diagnosis offered by the researchers is that the earlier approach ignored geometry. Because the two cameras observe the same target from different vantage points, the target’s image coordinates in one view bear a systematic, slowly varying relationship to its coordinates in the other. When one view’s tracker drifts or loses the target, its search region—the crop of the image that the tracker examines in the next frame—becomes spatially misaligned with the other view’s search region, and feeding misaligned regions into a cross-attention module compounds the error. Correct correspondence across viewpoints, the team argues, is a prerequisite for meaningful cross-view collaboration, and it is precisely this aspect that prior work left unaddressed.

GRPT closes that gap by treating the cross-view geometry itself as a time series forecasting problem. For each frame, the tracker produces bounding boxes in both the UAV and ground views, each defined by coordinates plus width and height. The researchers first compute the center point of each bounding box, then define the offset as the vector difference between the ground-view center and the UAV-view center. Over the past k frames, this yields an offset sequence that captures how the geometric relationship between the two perspectives evolves as the target moves and the cameras pan. This sequence is embedded into a high-dimensional feature space by a learned linear projection, augmented with the sinusoidal positional encodings of the original Transformer architecture to preserve temporal order, and fed into a stack of six transformer encoder layers. The decoder, also six layers deep with eight attention heads, autoregressively predicts the offset for the current frame through cross-attention against the encoded history and self-attention over its own previous predictions. Within each attention module, queries are compared against keys through a scaled dot product, and the resulting weights are applied to values, giving the network its capacity to model nonlinear temporal variation in the offset trajectory—something earlier recurrent approaches using LSTMs could capture only with far weaker long-range modeling.

Prediction alone, however, is not enough; the offset must be acted upon. The team’s second contribution, the Search Region Adjustment strategy, exploits the fact that a tracker’s response or confidence score reflects, to a useful degree, the quality of its tracking result. In each frame, the system compares the response scores of the UAV-view and ground-view search regions. The region with the higher score is judged to be of better quality and trusted as a reliable anchor; the center of the other region is then corrected by adding or subtracting the predicted offset. Concretely, when the ground view scores higher, the UAV-view search center is repositioned to the ground center minus the predicted offset, and vice versa when the UAV view scores higher. The two realigned search regions, along with the target template images, are then passed into the MvCL-TransT multi-view tracker, ensuring that even when one view is degraded by occlusion, out-of-view events, or optical interference, the geometric bridge reconstructed by GRPT pulls its search region back onto the target. The researchers observed that when the baseline tracker lost the target in one view, the discrepancy between its search region offset and the true offset grew dramatically—a failure mode their predicted offsets largely eliminated.

A further technical wrinkle arose during training. The offset predictor and the tracker are tightly coupled at inference: past tracking results feed the predictor, and predicted offsets reshape the search regions that determine future tracking results. Training the predictor only on ground-truth offset sequences creates a mismatch with the noisy, self-generated inputs it encounters at test time. To resolve this, the authors designed an iterative optimization procedure. The predictor is first trained on ground-truth offsets to yield an initial model, which is then embedded into the full tracking system to run over the training set; the resulting tracking outputs are converted into new offset sequences that retrain the predictor. Experiments showed that one round of this optimization was sufficient, after which further iterations produced negligible gains. Ablation studies confirmed the importance of each component: adding offset-based adjustment alone helped somewhat, adding the full SRA strategy helped more, and incorporating the optimized training delivered the largest share of the improvement, contributing up to 3.3 percent in precision gains in the UAV view beyond the unoptimized variant.

The evaluation was conducted on UGVT, a dataset purpose-built for the UAV-ground tracking task. It contains 210 pairs of synchronized video sequences—45 reserved for testing—spanning 21 target categories, more than 204,000 frames, over 20 distinct scenarios, and 10 challenging attributes, with ground footage captured from 1 to 4 meters and aerial footage from 10 to 50 meters. Against ten leading single-view trackers from 2019 to 2023, including DiMP, TransT, STARK, OSTrack, AiATrack, SparseTT, and CTTrack, as well as the multi-view MvCL-TransT baseline, GRPTrcker ranked first across success, precision, and normalized precision metrics in both views. The advantage widened on attribute-specific tests: in the ground view, gains over the baseline reached 3.7 percent on fast motion and 5.8 percent on motion blur sequences, while in the UAV view the method improved by 5.6 percent on low-resolution sequences. The researchers attribute this resilience to the offset signal itself, which is less susceptible to appearance-based confusions than the tracker’s similarity computations—when the ground view loses the target, the UAV view’s position plus the predicted offset can relocate it, and the reverse holds equally.

Qualitative results reinforce the quantitative story. In one sequence set against a cluttered background riddled with optical interference and lookalike objects, the baseline tracker’s poorly placed search region captured a wrong target, while GRPTrcker centered its adjusted region on the correct one. In another, a cyclist momentarily hidden behind a tree caused the baseline to generate a search region devoid of the target; the new method recovered it via the cross-view offset. Similar recoveries occurred when a running pedestrian briefly left the frame and when a high-altitude UAV shot degraded the target to low resolution amid similar distractors. In a basketball sequence combining occlusion and motion blur across the two views, and a second sequence featuring background clutter with the target drifting toward a distractor after reappearing, GRPTrcker maintained stable tracking where the baseline lost the trail.

The authors are candid about the framework’s principal limitation: speed. The baseline multi-view tracker ran at 20 frames per second, and GRPTrcker manages only 4 frames per second, a penalty incurred because the full-size transformer predictor must run every frame, on top of the inherent cost of processing two views. They point to lightweight prediction models and early-exit strategies as the most promising routes to real-time performance, and they see applications extending well beyond the benchmark, including autonomous driving, multi-camera surveillance networks, and fire rescue missions where aerial and ground robots must cooperatively follow people or hazards.

Conceptually, the study signals a shift in how multi-view tracking systems might be designed. Rather than treating geometric correspondence as an implicit byproduct of feature matching, GRPT makes it an explicit, learnable, temporally modeled quantity—and shows that the humble coordinate offset between two cameras carries enough signal to rescue a tracker precisely where appearance-based reasoning collapses. As drone-ground teams of machines become standard in inspection, security, and disaster response, frameworks that let different vantage points reason about their shared geometry may prove as important as the trackers themselves.