Airport scene multi-aircraft target tracking method and system based on motion structure modeling and semantic contrast learning

Abstract

The application relates to the technical field of computer vision, and particularly discloses an airport scene multi-aircraft target tracking method and system based on motion structure modeling and semantic contrast learning. The application takes MOTIP as a baseline, introduces a position structure perception auxiliary branch to explicitly encode six-dimensional motion geometric information such as speed, direction and scale change from historical trajectories, and injects a Transformer main body in the form of a residual through a multi-layer perception machine to endow the model with an internal modeling capability for trajectory dynamics, effectively improving the target discrimination degree under the condition of similar appearances; a grouping perception ID decoder is introduced to bias and modulate the query vector through a learnable scale group token, so that the attention mechanism can adapt to targets of different scales, and the identity confusion problem caused by the scale difference of targets near and far is significantly alleviated; and a cross-frame semantic contrast loss function is introduced to use the InfoNCE form to shorten the distance of the same target in adjacent frames and push away different targets at the feature level, thereby strengthening the consistency of the identity in the tracking process.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically relating to a method and system for tracking multiple aircraft targets in airport scenes based on motion structure modeling and semantic contrastive learning. Background Technology

[0002] With the development of the global air transport industry, the density of aircraft operations at airports continues to rise, placing higher demands on airport safety and efficiency in aircraft scheduling on runways and taxiways. Utilizing video sequences to automatically obtain aircraft spatial trajectories not only provides accurate data support for core operations such as conflict detection, taxiway path optimization, and abnormal behavior early warning, but also significantly reduces the manpower costs of manual monitoring. Therefore, highly robust multi-aircraft tracking technology has become one of the key issues in the construction of smart airports.

[0003] However, compared to typical pedestrian or vehicle scenarios, multi-target tracking at airports faces three more challenging issues: First, large passenger aircraft and ground vehicles coexist in the same field of view, resulting in a vast range of target sizes; second, aircraft paint schemes and tail markings are highly similar, and appearance drift is easily observed under varying lighting conditions, weather changes, and heat waves; third, aircraft frequently experience prolonged occlusion during towing, taxiing, and sharp turns, making traditional appearance-based target re-identification (Re-ID) models prone to identity switching. Figure 1 As shown, these typical scenarios demonstrate the challenges faced in multi-target tracking at airports, including multi-scale targets, targets with similar appearances, and severe occlusion.

[0004] Early tracking efforts often employed a two-stage "detection-association" framework. This involved generating candidate bounding boxes in each frame, then assembling cross-frame detections into trajectories using spatiotemporal consistency, and finally extending the target's trajectory through Kalman filters and IoU matching. Methods like SORT demonstrated excellent real-time performance but couldn't address identity fragmentation caused by severe occlusion. Subsequently, DeepSORT and BoT-SORT introduced deep Re-ID features in the association stage, improving occlusion recovery capabilities, but still experiencing mismatches when appearances were similar and scale differences were significant. OC-SORT and ByteTrack enhanced detection noise robustness through IoU frame interpolation and dual-threshold screening, but still lacked explicit modeling of target motion geometry. On the other hand, "integrated detection-tracking" models like FairMOT and JDE embedded detection and Re-ID together in the backbone network, significantly reducing inference latency, but limited sensitivity for multi-scale aircraft detection due to fixed anchor benchmarks. In the past two years, Transformer-based MOT frameworks (TransTrack, MOTR, MOTRv2, etc.) have emerged, which have the advantage of end-to-end decoupling and association due to their self-attention global modeling capabilities. However, most of them only use visual features to complete identity prediction, do not explicitly constrain cross-frame semantic consistency, and do not design special mechanisms for scale changes.

[0005] In summary, current research can be broadly categorized into three paradigms: post-detection tracking (TBD), joint detection tracking (JDT), and end-to-end query-based Transformer architecture (TBQ). For example... Figure 2 As shown, TBD decouples detection and association, making it easy to implement but lacking robustness to occlusion and appearance drift; JDT performs detection and Re-ID simultaneously within a shared feature framework, balancing real-time performance and accuracy, but is limited by scale adaptation capabilities; TBQ, on the other hand, achieves end-to-end unified modeling by introducing learnable trajectory queries and a global attention mechanism, demonstrating strong cross-frame association potential. Summary of the Invention

[0006] To address the shortcomings and problems of existing Transformer-based multi-target tracking methods (especially MOTIP) in complex airport scenarios, such as over-reliance on appearance information, neglect of motion geometry information, and sensitivity to target scale changes, which leads to easy identity confusion when targets are occluded or scales are mixed, this invention provides a multi-aircraft target tracking method and system for airport scenes based on motion structure modeling and semantic contrastive learning.

[0007] A method for multi-aircraft target tracking in airport environments based on motion structure modeling and semantic contrastive learning includes the following steps: Step 1: Obtain the video frame sequence and use a Transformer-based backbone network to extract features and detect targets in the current frame, obtaining the detection boxes of multiple targets in the current frame and their corresponding spatial feature embeddings. Step 2: Construct a location structure auxiliary branch. For each target detection box, construct a motion geometric feature vector based on its historical trajectory box, and map the motion geometric feature vector to a location structure feature. Then, fuse the location structure feature with the spatial feature embedding to obtain the fused feature embedding. Step 3: Construct a group-aware ID decoder, generate a scale-guided vector based on the detection box scale information of each target, and use the scale-guided vector to modulate the fused feature embedding to obtain a feature embedding with scale awareness capability. Step 4: Embed the scale-aware features as a query, perform attention interaction with the pre-stored historical trajectory features, and input the interaction result into the ID classification head to predict the identity ID corresponding to each target; Step 5: Construct the total loss function, which includes detection loss, ID classification loss, and cross-frame semantic comparison loss. Based on the total loss function, perform end-to-end joint optimization training on the multi-target tracking model to achieve multi-target tracking in complex scenes.

[0008] In the above-mentioned airport scene multi-aircraft target tracking method based on motion structure modeling and semantic contrastive learning, the motion geometric feature vector in step two includes one or more of the following: current target center point position, historical target center point position, normalized velocity vector, velocity magnitude, velocity direction, inter-frame distance of center position, and normalized frame interval.

[0009] In the above-mentioned airport surface multi-aircraft target tracking method based on motion structure modeling and semantic contrastive learning, step two involves fusing the positional structure features with the spatial feature embedding to obtain a fused feature embedding. Specifically, the positional structure features are injected as residuals into the spatial feature embedding through element-wise addition.

[0010] In the above-mentioned airport scene multi-aircraft target tracking method based on motion structure modeling and semantic contrastive learning, in step three, the scale guidance vector is generated according to the normalized detection box area of ​​the target. Specifically, the normalized detection box area is mapped through a linear transformation layer to obtain a bias vector consistent with the embedding dimension of the fused feature.

[0011] In the above-mentioned airport scene multi-aircraft target tracking method based on motion structure modeling and semantic contrastive learning, step three involves modulating the fused feature embedding using the scale-guided vector, specifically by adding the bias vector to the fused feature embedding.

[0012] The above-mentioned airport scene multi-aircraft target tracking method based on motion structure modeling and semantic contrastive learning further includes, in step three: dividing the targets into different scale groups according to a preset area threshold, and assigning different learnable group tokens as scale guidance vectors to targets in different groups.

[0013] The above-mentioned airport scene multi-aircraft target tracking method based on motion structure modeling and semantic contrastive learning, the cross-frame semantic contrastive loss in step five is constructed based on the InfoNCE loss function. Specifically, for the same target, two feature representations in two adjacent frames constitute a positive sample pair, and for feature representations of different targets or different frames in the same batch, a negative sample pair is constituted, so as to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs.

[0014] The above-mentioned multi-aircraft target tracking method for airport surfaces based on motion structure modeling and semantic contrastive learning has the following total loss function:

[0015]

[0016]

[0017]

[0018] In the formula: This represents the supervisory loss in identity recognition; Indicates detection loss; Indicates cross-frame semantic contrast loss; and This is the loss weight hyperparameter, used to balance the impact of each loss term.

[0019] This invention also provides a multi-aircraft target tracking system for airport surfaces based on motion structure modeling and semantic contrastive learning, comprising: The target detection module is used to acquire video frame sequences and use a Transformer-based backbone network to extract features and detect targets in the current frame, obtaining detection boxes of multiple targets in the current frame and their corresponding spatial feature embeddings. The motion geometry fusion module is used to construct a position structure auxiliary branch. For each target detection box, a motion geometry feature vector is constructed based on its historical trajectory box, and the motion geometry feature vector is mapped to a position structure feature. The position structure feature is fused with the spatial feature embedding to obtain the fused feature embedding. The scale-aware modulation module is used to construct a group-aware ID decoder, generate a scale-guided vector based on the detection box scale information of each target, and use the scale-guided vector to modulate the fused feature embedding to obtain a feature embedding with scale awareness capability. The identity decoding module is used to embed the scale-aware features as a query, perform attention interaction with the pre-stored historical trajectory features, and input the interaction result into the ID classification head to predict the identity ID corresponding to each target. The joint training optimization module is used to construct the total loss function, which includes detection loss, ID classification loss and cross-frame semantic comparison loss. Based on the total loss function, the multi-object tracking model is jointly optimized and trained end-to-end to achieve multi-object tracking in complex scenes.

[0020] Compared with the prior art, the beneficial effects of the present invention are: This invention enhances the Transformer-based MOTIP model with a series of improvements, including a motion structure-assisted branch for explicit motion modeling, a group-aware ID decoder for scale-adaptive attention, and a cross-frame semantic contrastive loss to ensure temporal feature consistency. By explicitly fusing velocity, orientation, and scale information, and adapting group tokens to multi-scale targets, and by using contrastive learning to constrain cross-frame consistency, these three elements work synergistically to effectively address the three core challenges of scale variation, appearance similarity, and occlusion. While maintaining the end-to-end advantages and real-time performance of Transformer, it significantly improves occlusion recovery and scale robustness in airport scenarios, thereby enhancing performance in complex airport ground target tracking.

[0021] This invention introduces a location-structure auxiliary branch based on the original MOTIP baseline. By explicitly modeling the target's motion trajectory and scale changes, this branch reuses the Hungarian matching results from the baseline model MOTIP without introducing an additional spatiotemporal matching module. This achieves "implicit alignment" within the tensor dimension through the established one-to-one correspondence between the detection boxes and the real targets. Thus, it provides powerful spatiotemporal auxiliary features for subsequent identity decoding without introducing additional computational overhead. By explicitly modeling structural location information, it guides visual features to learn the behavioral differences of targets in the temporal dimension, assisting visual features in identity prediction in multi-target tracking tasks. To ensure the stability of cross-frame information, the branch maintains two cache tensors, used to store the historical target boxes and their corresponding frame numbers from the previous forward propagation. This design enables the model to continuously perceive the spatiotemporal evolution patterns of targets, thereby achieving more accurate identity discrimination in similar appearance scenes and improving the ability to distinguish similar-looking targets.

[0022] The group-aware ID decoder introduced in this invention maintains the same structure as the original MOTIP model without introducing additional decoding paths. Instead, it introduces a "group-aware" prior control mechanism in the input feature processing stage, using the scale information of the target to guide the attention mechanism of the decoder, making it more scale-aware. This improves the ability to distinguish targets of different sizes and significantly enhances the model's robustness in recognizing scale-sensitive targets. In particular, it demonstrates stronger robustness and generalization ability in practical applications where small targets frequently appear in airport scenarios.

[0023] This invention introduces cross-frame semantic contrast loss in addition to detection loss and classification loss when constructing the loss function to enhance the consistency of semantic representation of the same target in different time frames. It does not require additional labels and is based solely on existing matching information (GT ID) and frame sequences. It achieves efficient computation through simple tensor rearrangement and has good scalability and compatibility. By maintaining the consistency of representation of the same target in adjacent frames, it significantly improves the ability to distinguish similar targets in tracking tasks, effectively alleviates the ID switch phenomenon, and complements the standard loss to build a more robust tracking model. Attached Figure Description

[0024] Figure 1 For multi-target tracking scenarios at airports in different settings; Figure 2 This is a schematic diagram of the existing TBD, JDT, and TBQ systems. Figure 3 This is a schematic diagram of the overall system structure of the present invention; Figure 4 This is a schematic diagram of the auxiliary branch module structure of the present invention. Figure 5 Qualitative comparison results of different methods in clear sky scenes with interwoven multi-scale and occlusion; Figure 6 Qualitative comparison results of different methods under rainstorm scenarios with drastic changes in lighting and weather; Figure 7 Visualize the t-SNE distribution of trajectory features under different improvement strategies in multi-target tracking. Detailed Implementation

[0025] Currently, the main technologies in MOT research include the following: (1) Application of traditional MOT and Transformer based on detection-association in MOT Multi-object tracking (MOT) research initially followed the "tracking-by-detection" paradigm, which first generates candidate boxes in each frame and then assembles cross-frame detections into a trajectory through spatiotemporal consistency. SORT uses uniform Kalman filtering and Hungarian matching to complete data association, achieving extremely high inference speed, but it treats occlusion as target disappearance, easily causing IDSwitch. DeepSORT introduces a CNN-based Re-ID descriptor to significantly reduce false matches caused by short-term occlusion, but it still fails when appearances are extremely similar or scale changes drastically. ByteTrack and OC-SORT preserve low-scoring detections and IoU trajectory frame completion through "dual thresholding," enhancing robustness against missed detections, but they do not explicitly model appearance drift. BoT-SORT combines high-precision Re-ID and motion consistency, improving the benchmark performance of MOT17, but often requires cumbersome threshold tuning in large-scale span scenarios. Overall, the limitations of traditional detection-association frameworks are: 1. Motion models and appearance models are independent of each other, making it difficult to form a globally optimal joint optimization; 2. They rely on heuristic thresholds and are sensitive to occlusion intensity and target scale; 3. Cross-frame association only uses local information and cannot fully capture long-range dependencies.

[0026] Self-attention can build global dependencies across entire frames or even entire sequences, providing a new approach for end-to-end tracking. TrackFormer was the first to extend the DETR decoder to a temporally recursive structure, using the "seed" query from the previous frame to associate the target in the next frame, achieving unified detection and association; however, its recursion length is limited, and it lacks explicit memory for occluded targets. MOTR adds cross-frame cross-attention at the backbone level while maintaining learnable trajectory queries, significantly improving long-range association capabilities; however, because the decoder directly processes all historical queries, memory usage is proportional to frame length. MOTRv2 alleviates the memory bottleneck by using sparse updates and decoupled query strategies, but still mainly relies on appearance features for identity determination. MOTIP adds an ID decoder and dynamic dictionary update mechanism to the multi-scale attention of Deformable-DETR, balancing appearance changes and real-time performance; however, it has not yet explicitly modeled motion geometry and scale differences, and false associations still occur when there are mixed distant and near aircraft and long-term occlusion.

[0027] (2) Application of motion priors and contrastive learning in video understanding and tracking In the field of video understanding, motion priors have long been used to compensate for the lack of appearance information in a single frame. Traditional methods often rely on optical flow or Kalman / uniform velocity models to describe local motion consistency; frameworks such as FlowTrack concatenate optical flow features with CNN features to improve single-target tracking, while LSTR and TransTrack explicitly inject position encoding or velocity embedding into the Transformer, improving the reliability of target association in fast-moving and short-term occlusion scenarios. However, these motion priors are usually limited to short-term relative displacement, and their modeling of long-range trajectory morphology (such as velocity trends, direction changes, and scale evolution) remains coarse. Furthermore, they are often "post-fused" with appearance branches, making it difficult to achieve end-to-end unified learning.

[0028] Meanwhile, contrastive learning has achieved breakthroughs in unsupervised video representation with its InfoNCE loss: methods such as SimCLR-V and MoCoV2-T generate positive and negative sample pairs through frame-level augmentation, encouraging models to capture temporal consistency; VID-CLR, SCE-Net, and others further combine spatial location or motion boundaries to achieve finer spatiotemporal alignment. In the direction of target tracking, works such as UniTrack and MOTContrast attempt to apply contrastive constraints to the embedding of the same trajectory at different times to alleviate appearance drift. However, existing approaches either rely on large-capacity memory queues to increase the size of negative samples, bringing additional GPU memory and synchronization overhead; or they are only trained in offline feature space, which provides limited help to the online association process.

[0029] (3) Application of multi-target tracking algorithm in airports Airports involve the high-density collaborative operation of aircraft, support vehicles, and special equipment. Real-time, multi-target visual perception is a key component of apron safety monitoring, ground traffic management, and autonomous taxiing systems. Early solutions relied heavily on radar echoes or ADS-B broadcasts for macro-level situational awareness, but these solutions suffered from insufficient resolution and severe occlusion shadows in close-range apron scenarios, making it difficult to support fine-grained management at the gate level. This has spurred research into video-based visual multi-target tracking.

[0030] Traditional CNN+Kalman / SORT systems often suffer from mismatches and frequent ID drift due to the unique characteristics of airports, such as differences in target scale, large areas of low-texture background, and appearance degradation caused by high reflectivity / backlighting. Recent works such as A-STRM have improved nighttime tracking accuracy by introducing terminal overhead thermal imaging, but still rely on external sensor switching. AirMOT, on the other hand, has achieved high MOTA on the publicly available Apron-MOT dataset by using a detection-association framework based on DINO-DETR, but it has a high false negative rate for small targets at a distance and is not well adapted to changes in aircraft scale during rapid turns / pushback phases.

[0031] A comprehensive evaluation of existing methods reveals that: motion priors mostly rely on the assumption of uniform velocity or simple Kalman filtering, lacking robustness to complex sequences such as taxiing-turning-stopping-traction; appearance representation is affected by similar aircraft paint schemes and lighting glare, making it difficult for pure visual embedding to maintain long-term consistency; in semi-enclosed scenarios like airports, target trajectories have obvious road constraints and physical reachability zones, but current Transformer-MOT frameworks rarely explicitly incorporate "flight zone motion patterns" into the decoding process.

[0032] To address the aforementioned pain points, this invention proposes an improved version of MOTIP that combines positional structure-assisted branching with semantic contrast constraints. This version enhances motion representation using the geometric consistency of "historical trajectory - current detection" without relying on additional sensors; suppresses appearance degradation through cross-frame contrastive learning; and maintains end-to-end inference speed. This provides a new solution for multi-aircraft tracking on airport tarmac that balances accuracy and real-time performance. The invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0033] Example 1: This example provides a multi-aircraft target tracking method for airport surfaces based on motion structure modeling and semantic contrastive learning, as detailed below.

[0034] This embodiment uses MOTIP as the baseline model. MOTIP is a multi-object tracking method based on an end-to-end Transformer architecture. Its core idea is to transform the multi-object tracking problem into a frame-level identity (ID) prediction problem. Compared to the traditional two-stage "detection-association" framework, MOTIP directly predicts the identity of the target while detecting it, thus achieving joint optimization of detection and identity assignment. MOTIP uses Deformable DETR as its basic network architecture. For a given video frame sequence input... First, Deformable DETR is used to extract spatial feature embeddings and corresponding target location bounding boxes from each frame, and then outputs a fixed-length bounding box. Candidate target set:

[0035] in, For bounding box coordinates (using) Format), Embed a representation for the corresponding target features.

[0036] Considering the challenging nature of airport scene target tracking tasks—characterized by complex scenarios, significant differences in target scale, and high similarity in appearance—relying solely on visual features is insufficient for stable and reliable target identification. This invention, building upon the existing MOTIP baseline, introduces a Motion-Structure Branch. By explicitly modeling the target's motion trajectory and scale changes, this branch assists visual features in identity prediction within multi-target tracking tasks. Without introducing an additional spatiotemporal matching module, this branch reuses the Hungarian matching results from the baseline model MOTIP, achieving "implicit alignment" within the tensor dimension through the established one-to-one correspondence between detection boxes and real targets. This provides powerful spatiotemporal auxiliary features for subsequent identity decoding without introducing additional computational overhead.

[0037] like Figure 4 As shown, this location-structure auxiliary branch directly extracts motion and structural features from historical trajectories, avoiding redundant calculations for the target in the current frame. For the current frame... The Middle Detection box and its historical trajectory frame We construct the following 6-dimensional geometric structural features. : Current target center point location:

[0038] Location of the historical target center point:

[0039] Velocity vector (normalized center position difference):

[0040] Speed ​​magnitude (module length):

[0041] Velocity direction (angle):

[0042] Inter-frame distance at center position:

[0043] Normalized frame interval:

[0044] Finally, the positional structure feature vector is:

[0045] This 6-dimensional feature in frame interval If a historical bounding box is missing, it will be set to zero to avoid abnormal information interfering with model learning.

[0046] Subsequently, the positional structure feature vector The data is fed into a two-layer MLP network and mapped to 256-dimensional embedding features consistent with the DETR output. ,get

[0047] Finally, the positional structural features are embedded with the target features output by the DETR in the current frame through element-wise addition. To merge:

[0048] This fusion strategy guides visual feature learning to discriminate targets across temporal dimensions by explicitly modeling structural location information, thereby improving the ability to distinguish similar-looking targets. To ensure the stability of cross-frame information, two cache tensors are maintained within each branch, used to store the historical target bounding boxes and their corresponding frame numbers from the previous forward propagation. This design enables the model to continuously perceive the spatiotemporal evolution patterns of targets, thus achieving more accurate identity recognition in similar-looking scenes.

[0049] In airport scenarios, aircraft targets exhibit significant size variations. Small targets, due to their low resolution and insufficient appearance information, often lead to incorrect ID decoder assignment. To address this, this paper proposes a group-aware ID decoder, aiming to leverage target scale information to guide the decoder's attention mechanism, making it more "scale-aware" and thus improving its ability to distinguish targets of different sizes. This decoder maintains structural consistency with the original MOTIP model, without introducing additional decoding paths. Instead, it incorporates a "group-aware" prior control mechanism during the input feature processing stage.

[0050] Specifically, for each target, its corresponding DETR output embedding Through an additional scale guide vector Adjusting attention or representation processes, in which Derived from its normalized bounding box area size:

[0051] in, and The target Width and height, Let be the image area. To enhance the expressiveness of the scale signal, we will... The vector is fed into a linear layer for transformation and then used as a bias vector in the target embedding.

[0052] In practical implementations, if dimension-consistent embedding is used, for example... Then the guide item can be directly connected with... Adding them together creates features with scale perception capabilities. These adjusted features are then fed into a unified ID decoder for classification prediction, eliminating the need to structurally distinguish between "small target paths" and "large target paths," thus maintaining parameter sharing and computational efficiency within the network. Furthermore, to further enhance the grouping guidance effect, the model checks whether the area of ​​each target exceeds a preset threshold in the forward function. Different attentional guidance is applied to them:

[0053] Samples from different groups share the same structure but can use different group embeddings (e.g., two types of embedding vectors) as guide vectors. This design introduces a dynamic perception mechanism without increasing the model parameters. Finally, all group samples are merged and fed into the ID classifier for standard identity prediction and loss calculation.

[0054] The group-aware ID decoder used in this invention constructs structurally consistent and semantically distinguishable feature representations for multi-scale targets through a scale prior guidance mechanism, which significantly improves the model's robustness in recognizing scale-sensitive targets. In particular, it demonstrates stronger robustness and generalization ability in practical applications where small targets frequently appear in airport scenarios.

[0055] In multi-object tracking (MOT) tasks, the identity prediction module needs to be able to distinguish between different objects. To this end, the baseline model MOTIP transforms object association in MOT into an end-to-end learnable classification problem through contextual ID prediction, using standard cross-entropy loss to supervise the training process of the ID decoder. Specifically: let the target embedding features output by the ID decoder be... After transformation by the classifier, the predicted logits of the target identity are: ,in The total number of identity categories, and the target's real ID tag. The supervised loss for identity recognition is defined using the cross-entropy form as:

[0056] On the other hand, MOTIP's detection module based on the DETR architecture employs the following two losses to jointly optimize the target detection task: Bounding box regression loss: using Loss and GIoU loss.

[0057] Classification loss: Focal Loss is used to suppress class imbalance.

[0058] If the detection loss is expressed as Its general form is:

[0059] While the aforementioned loss functions are effective in detection and classification tasks, inconsistencies in semantic features across frames remain, especially when the same target experiences pose changes, occlusion, or resolution fluctuations in consecutive frames, potentially causing its embedded features to shift in the semantic space. To address this, this invention introduces Cross-frame Semantic Contrastive Loss (CSCL) to enhance the consistency of semantic representation of the same target across different time frames. This method is based on InfoNCE and uses... Cosine similarity is calculated using the normalized embedding vectors. For the same target... In adjacent frames and Feature representation in and The comparative losses are as follows:

[0060] in For temperature parameters, This represents the number of samples in the same batch. Only adjacent frames with Δt = 1 are selected as positive sample pairs, and targets with other different IDs constitute negative samples. All embedded features are output by the ID decoder and processed... Normalization. This loss function requires no additional labels; it is based solely on existing matching information (GT ID) and frame sequences, achieving efficient computation through simple tensor rearrangement, and possesses good scalability and compatibility. In practical implementation, the CSCL loss is integrated into the ID decoder training process, forming a joint optimization strategy of "discrimination + contrast," which significantly improves the consistency of target representation.

[0061] Finally, the overall loss function used in this paper is defined as follows:

[0062] in, and This is the loss weight hyperparameter, used to balance the impact of each loss term.

[0063] Cross-frame semantic contrast loss significantly improves the ability to distinguish similar-looking targets in tracking tasks by maintaining the consistency of representation of the same target in adjacent frames, effectively alleviating the ID switch phenomenon, and complementing the standard loss to build a more robust tracking model.

[0064] Example 2: An airport surface multi-aircraft target tracking system based on motion structure modeling and semantic contrastive learning, used to execute the method of Example 1, including: Target detection module: Used to acquire video frame sequences and use a Transformer-based backbone network to extract features and detect targets in the current frame, obtaining detection boxes of multiple targets in the current frame and their corresponding first target feature embeddings; Motion geometry fusion module: Used to construct position structure auxiliary branches. For each target, a motion geometry feature vector is constructed based on its historical trajectory information, and the motion geometry feature vector is mapped to a motion geometry embedding. The motion geometry embedding is fused with the first target feature embedding to obtain a second target feature embedding. Scale-aware modulation module: Used to construct a group-aware ID decoder, generate a scale-guided vector based on the detection box scale information of each target, and use the scale-guided vector to modulate the second target feature embedding to obtain the third target feature embedding; Identity decoding module: used to embed the third target feature as a query, perform attention interaction with the pre-stored historical trajectory feature memory, and input the interaction result into the ID classification head to predict the identity ID corresponding to each target; Joint Training Optimization Module: Used to construct the total loss function, which includes detection loss, ID classification loss, and cross-frame semantic contrast loss, and to perform end-to-end joint optimization training on the multi-target tracking model based on the total loss function; wherein, the cross-frame semantic contrast loss is used to bring the feature representations of the same target in adjacent frames closer together and push away the feature representations of different targets.

[0065] Experimental Example: To comprehensively evaluate the practical application capability of the multi-target tracking method proposed in this invention in airport scenarios, we conducted a systematic experimental study on the AGVS-T22 (Airport Ground Video Surveillance-T22) dataset. The AGVS-T22 dataset is collected from real airport ground surveillance videos and covers a variety of typical complex situations, including target scale variations, frequent occlusion between targets and with background elements, strong lighting changes (such as front lighting, backlighting, and nighttime scenes), various weather conditions (such as rain, sunshine, and fog / haze), and dynamic camera movement. These factors collectively constitute the practical challenges of target tracking in airport scenarios and provide a realistic and comprehensive experimental environment for this research.

[0066] Specifically, the training set of this dataset contains 37 videos, totaling 93,041 frames, while the test set contains 14 videos, totaling 25,095 frames. Each frame in the dataset provides accurate target annotation information, including bounding box locations and corresponding target IDs, providing a high-quality foundation for the training and evaluation of multi-target detection and tracking models.

[0067] In terms of evaluation metrics, this paper adopts the mainstream evaluation system in the field of multi-object tracking, comprehensively examining the model's performance in terms of detection accuracy, identity consistency, and global tracking performance. The multi-object tracking accuracy (MOTA) comprehensively considers false detections, false negatives, and ID switching, and its definition is as follows:

[0068] Among them, FN t FPt These represent the number of missed detections and the number of false detections in frame t, respectively. IDSW t GTt represents the number of identity switching, and GTt represents the number of real targets in this frame.

[0069] Identity retention capability is measured by the IDF1 (Identification F1 Score), which is defined as:

[0070] Here, DetA represents detection accuracy, and AssA, through joint analysis of these metrics, allows us to comprehensively evaluate the model's multi-target tracking performance from multiple dimensions, including detection, identity preservation, association accuracy, and global robustness. HOTA represents target association accuracy, providing a more balanced reflection of overall tracking performance.

[0071] In implementation, we chose the Deformable DETR with a ResNet-50 backbone as our default DETR detector because it is a common choice for downstream tasks. Similar to previous work, this invention also utilizes weights pre-trained on the COCO dataset as initialization. Relative position encoding is applied in the ID decoder because tracking focuses more on relative temporal relationships than absolute timestamps. To minimize unnecessary additional modules, the hidden dimension throughout the model is kept consistent with the Deformable DETR, i.e., C = 256. Since the ID dictionary can be reused, it is only necessary to ensure that K is not less than the maximum number of targets per frame. Here, for simplicity, we set K to 50. The deep learning framework used during training is PyTorch, and the hardware configuration is an NTX 3090 GPU. Adam is used as the gradient optimizer, and the batch size is set to 1. In our experiments, the supervised weight coefficients λcls, λL1, λgiou, and λid are set to 2.0, 5.0, 2.0, and 1.0, respectively, and λcscl is set to 0.1. The inference thresholds λdet, λnew, and λid are set to 0.3, 0.6, and 0.2, respectively.

[0072] To validate the performance of the proposed method in multi-target tracking scenarios in complex airport environments, we systematically compared it with several mainstream benchmark methods on the AGVS-T22 test set. These included SORT based on Kalman filtering, DeepSORT, OC-SORT, and ByteTrack which fuse appearance features, as well as JDE, FairMOT, and CenterTrack which combine pedestrian detection and ReID tasks. Furthermore, we also evaluated recent high-performing multi-target tracking methods based on the Transformer architecture.

[0073]

[0074] As shown in Table 1, traditional two-stage frameworks (SORT, DeepSORT, OC-SORT, etc.) are generally limited in terms of MOTA and IDF1, especially DeepSORT, which only achieves 33.6% MOTA and 38.2% IDF1, revealing its inadequacy in handling large-scale changes and severe occlusion. In contrast, BoT-SORT and ByteTrack, based on Hungarian matching and backtracking optimization, improve MOTA to approximately 65%, but still have significant gaps in identity consistency (IDF1 approximately 56%).

[0075] The end-to-end methods MOTR and MOTRv2, by unifying detection and association modeling, achieve 69.7% and 84.3% respectively on MOTA, and also generate high IDF1 (82.9% / 88.2%), demonstrating the potential of the Transformer-based framework. Our proposed method maintains absolute leadership across all three metrics, reflected in MOTA 85.4%, MOTR 81.7%, and IDF1 88.5%; compared to the second-best MOTRv2, MOTA is improved by 1.1 percentage points, and IDF1 is further improved by 0.3 percentage points. Notably, when facing high-speed traffic and frequently occluded taxiway segments, our method significantly reduces misfollowing and ID swapping events, demonstrating the crucial role of motion-structure auxiliary branches and semantic contrast constraints in identity preservation. Overall, the experimental results fully demonstrate the robustness and effectiveness of the proposed improvement in complex airport scenarios.

[0076] Meanwhile, this experimental example tested the stepwise ablation results of different methods on the AGVS-T22 test set, as shown in Table 2 below.

[0077]

[0078] As shown in Table 2, the baseline model (pure MOTIP) laid a benchmark for subsequent improvements at the levels of MOTA 77.4%, IDF1 83.8%, and HOTA 68.1%, but still experienced 81 identity swaps (IDSW), revealing its limitation of insufficient identity consistency in scenarios with highly similar appearances. After introducing only the cross-frame semantic contrast loss LCSC, the detection accuracy remained largely unchanged (DetA 69.4% → 69.7%), while IDSW plummeted to 16 times (–80.2%), and MOTA simultaneously improved to 85.4%, indicating that this loss effectively enhanced the consistency and discriminative power of cross-frame features. Subsequently, applying the motion-structure auxiliary branch to historical trajectory features, without changing the detection end structure, improved IDF1 to 88.3% and HOTA to 71.9%, and further compressed IDSW to 24 times, verifying that explicit injection of motion geometry information significantly promotes identity stability. If a group-aware ID decoder is then added, the DetA and AssA metrics show a slight increase, while MOTP and HOTA also show gains, indicating that the robustness of group tokens to scale differences is mainly reflected in detection and localization accuracy. After combining these three improvements, the model of this invention achieves the highest results in MOTA (85.9%), IDF1 (88.9%), HOTA (73.2%), and MOTP (82.7%), while IDSW is compressed to only 9 times.

[0079] This experiment tested three time points in a multi-scale – shading-intertwined clear sky scenario and a rainstorm scenario with rapidly changing lighting and weather. , , The qualitative comparison yielded the following results: Figure 5 and Figure 6 As shown. In Figure 5 In the first frame, both MOTR and MOTIP showed missed detections and false detections in the fuselage / wing region at frame t+Δt1. At the same location, our method, leveraging the velocity and scale compensation provided by the motion-structure branch for DETR features, still outputs complete and accurate bounding boxes, demonstrating the significant improvement in recall rate for weakly textured targets by geometric priors. Meanwhile... Figure 6In the process, at time t+Δt2t+\Delta t_{2}t+Δt, approximately one thousand frames later, the MOTIP undergoes ID replacement, yet our method still maintains consistent trajectory identifiers. This is thanks to the historical embedding discrimination information injected into the Key / Value pairs by the group-aware ID decoder, and the consistency constraint of cross-frame semantic contrast loss on the embeddings of the same trajectory, making re-identification more stable and reliable during long-term tracking.

[0080] To further analyze the improvement effect, the 50 frames of trajectory were embedded and reduced to a two-dimensional plane using t-SNE. The results are as follows: Figure 7 As shown, (a) is the baseline MOTIP, with severe inter-class aliasing and obvious intra-class tailing; after adding the motion-structure branch, subgraph (b) shows that targets with large differences in motion vectors are effectively separated; when only cross-frame contrastive learning is applied, subgraph (c) shows compact intra-class targets, but some clusters are still close to each other; when both are combined with group decoding, subgraph (d) forms an ideal distribution of "single peak within the class and wide separation between classes". This process clearly reveals that geometric prior first widens the distance between targets with different motion patterns at the macro level, and contrastive learning then tightens the samples with the same trajectory at the micro level, ultimately shaping a more discriminative embedding space and providing more sufficient mutual information for subsequent matching.

[0081] Visual comparisons across the two scenarios demonstrate that this method can simultaneously improve detection completeness and identity consistency even under typical complex airport conditions such as drastic target scale changes, long temporal gaps or severe occlusion, and high degree of visual similarity. Combined with the quantitative results above, this further illustrates that motion geometry priors effectively compensate for detection gaps caused by insufficient appearance information, contrastive learning significantly reduces cross-temporal ID drift, and the fusion of these two approaches enables the model to maintain a stable and consistent trajectory even in ultra-long sequences (>1000 frames), providing more reliable target-level data support for airport operational situational awareness.

[0082] The above description is only a preferred embodiment of the present invention and does not limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for multi-aircraft target tracking on an airport surface based on motion structure modeling and semantic contrastive learning, characterized in that: Includes the following steps: Step 1: Obtain the video frame sequence and use a Transformer-based backbone network to extract features and detect targets in the current frame, obtaining the detection boxes of multiple targets in the current frame and their corresponding spatial feature embeddings. Step 2: Construct a location structure auxiliary branch. For each target detection box, construct a motion geometric feature vector based on its historical trajectory box, and map the motion geometric feature vector to a location structure feature. Then, fuse the location structure feature with the spatial feature embedding to obtain the fused feature embedding. Step 3: Construct a group-aware ID decoder, generate a scale-guided vector based on the detection box scale information of each target, and use the scale-guided vector to modulate the fused feature embedding to obtain a feature embedding with scale awareness capability. Step 4: Embed the scale-aware features as a query, perform attention interaction with the pre-stored historical trajectory features, and input the interaction result into the ID classification head to predict the identity ID corresponding to each target; Step 5: Construct the total loss function, which includes detection loss, ID classification loss, and cross-frame semantic comparison loss. Based on the total loss function, perform end-to-end joint optimization training on the multi-target tracking model to achieve multi-target tracking in complex scenes.

2. The airport surface multi-aircraft target tracking method based on motion structure modeling and semantic contrastive learning according to claim 1, characterized in that: In step two, the motion geometric feature vector includes one or more of the following: current target center point position, historical target center point position, normalized velocity vector, velocity magnitude, velocity direction, inter-frame distance of center position, and normalized frame interval.

3. The airport surface multi-aircraft target tracking method based on motion structure modeling and semantic contrastive learning according to claim 1, characterized in that: In step two, the positional structural features and the spatial feature embedding are fused to obtain the fused feature embedding. Specifically, the positional structural features are injected as residuals into the spatial feature embedding through element-wise addition.

4. The airport surface multi-aircraft target tracking method based on motion structure modeling and semantic contrastive learning according to claim 1, characterized in that: In step three, the scale-guided vector is generated based on the normalized detection box area of ​​the target. Specifically, the normalized detection box area is mapped through a linear transformation layer to obtain a bias vector consistent with the embedding dimension of the fused feature.

5. The airport surface multi-aircraft target tracking method based on motion structure modeling and semantic contrastive learning according to claim 4, characterized in that: In step three, the fusion feature embedding is modulated using the scale-guided vector, specifically by adding the bias vector to the fusion feature embedding.

6. The airport surface multi-aircraft target tracking method based on motion structure modeling and semantic contrastive learning according to claim 1, characterized in that: Step three further includes: dividing the target into different scale groups according to a preset area threshold, and assigning different learnable group tokens as scale guidance vectors to the targets in different groups.

7. The airport surface multi-aircraft target tracking method based on motion structure modeling and semantic contrastive learning according to claim 1, characterized in that: The cross-frame semantic comparison loss in step five is constructed based on the InfoNCE loss function. Specifically, for the same target, two feature representations in two adjacent frames constitute a positive sample pair, and for feature representations of different targets or different frames in the same batch, a negative sample pair is constituted, so as to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs.

8. The airport surface multi-aircraft target tracking method based on motion structure modeling and semantic contrastive learning according to claim 1, characterized in that: The total loss function is: ； ； ； ； In the formula: This represents the supervisory loss in identity recognition; Indicates detection loss; Indicates cross-frame semantic contrast loss; and This is the loss weight hyperparameter, used to balance the impact of each loss term.

9. A multi-aircraft target tracking system for airport surfaces based on motion structure modeling and semantic contrastive learning, characterized in that: include: The target detection module is used to acquire video frame sequences and use a Transformer-based backbone network to extract features and detect targets in the current frame, obtaining detection boxes of multiple targets in the current frame and their corresponding spatial feature embeddings. The motion geometry fusion module is used to construct a position structure auxiliary branch. For each target detection box, a motion geometry feature vector is constructed based on its historical trajectory box, and the motion geometry feature vector is mapped to a position structure feature. The position structure feature is fused with the spatial feature embedding to obtain the fused feature embedding. The scale-aware modulation module is used to construct a group-aware ID decoder, generate a scale-guided vector based on the detection box scale information of each target, and use the scale-guided vector to modulate the fused feature embedding to obtain a feature embedding with scale awareness capability. The identity decoding module is used to embed the scale-aware features as a query, perform attention interaction with the pre-stored historical trajectory features, and input the interaction result into the ID classification head to predict the identity ID corresponding to each target. The joint training optimization module is used to construct the total loss function, which includes detection loss, ID classification loss and cross-frame semantic comparison loss. Based on the total loss function, the multi-object tracking model is jointly optimized and trained end-to-end to achieve multi-object tracking in complex scenes.

Images