Track segmentation method, system and equipment based on space-time perception and adaptive learning
By employing a Transformer-based self-supervised pre-training and online learning approach, this study addresses the issues of high training costs and insufficient generalization ability in traditional deep learning for track segmentation. It achieves an efficient and interpretable track segmentation model suitable for air target identification and control.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA ACADEMY OF ELECTRONICS AND INFORMATION TECHNOLOGY OF CHINA ELECTRONICS TECHNOLOGY GROUP CORPORATION
- Filing Date
- 2025-12-25
- Publication Date
- 2026-04-21
AI Technical Summary
Traditional deep learning methods suffer from high training costs, insufficient generalization ability, and poor model interpretability in track segmentation.
We employ a Transformer-based encoder-decoder model for self-supervised pre-training, combining mask reconstruction and contrastive learning tasks. We utilize an unlabeled dataset for pre-training and adapt the model to specific tasks through supervised fine-tuning and online learning, while using a dynamic database for model fine-tuning.
It reduces training time costs, improves the model's generalization ability and interpretability, enables it to quickly adapt to new scenarios and environmental changes, and provides more reliable track segmentation support.
Smart Images
Figure CN121904095A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, system and device for track segmentation based on spatiotemporal perception and adaptive learning. Background Technology
[0002] Track segmentation refers to the technique of dividing an aircraft's track at inflection points or changes in motion, based on real-time or historical track data, using a specific segmentation method. For modern radar, target segmentation is one of its most fundamental tasks. Track segmentation can elevate raw track data into high-level knowledge that is interpretable and operable, thereby helping researchers perform target behavior recognition, intent understanding, and trajectory anomaly detection. It can also help understand the cooperative relationships between multiple targets, forming a unified traffic situation map. Due to the increasing complexity of radar detection environments, especially interference from strong clutter and the lack of relevant prior knowledge, the uncertainty of radar measurements has become extremely severe. Therefore, the problem of radar target track segmentation is gradually becoming a major research issue in radar data processing algorithms.
[0003] Current identification methods based on radar measurement data mainly employ deep learning-based track segmentation. These methods train machine learning models to achieve classification capabilities by constructing relevant sample datasets. This approach significantly reduces the model's reliance on prior knowledge, thereby improving its robustness and adaptability. Specifically, this method first labels the behaviors within each time period of the original track dataset, such as straight-ahead movement, circling, turning, and patrolling. Then, it preprocesses the track data, removing outliers and missing values to construct a dataset suitable for neural network training. Finally, it builds a deep learning model suitable for handling track segmentation problems and trains it using the track dataset. This approach significantly reduces the model's dependence on prior knowledge, thus improving its robustness and adaptability.
[0004] However, this approach also has certain limitations. Specifically, deep learning methods require large-scale, high-quality labeled datasets, which are extremely time-consuming; for complex and rare scenarios, there may be no labeled data, making it prone to overfitting to specific patterns in the training dataset, resulting in poor adaptability to unseen scenarios, and its generalization ability needs further improvement; deep learning models are usually "black box" models, with poor interpretability of the decision-making process. Summary of the Invention
[0005] The technical problem to be solved by this invention is to address the issues of high training cost, insufficient generalization ability and poor model interpretability in traditional deep learning methods, and to provide a trajectory segmentation method, system and device based on spatiotemporal awareness and adaptive learning.
[0006] The trajectory segmentation method based on spatiotemporal awareness and adaptive learning according to embodiments of the present invention includes: S1, preprocess the original track data to obtain unlabeled dataset and labeled dataset; S2, Construct a Transformer-based encoder-decoder model, and use the unlabeled dataset to perform self-supervised pre-training on the model through mask reconstruction and contrastive learning tasks to obtain a pre-trained model; S3. Based on the pre-trained model, add a segmentation decision head and perform supervised fine-tuning using the labeled dataset to obtain the track segmentation model. S4, Deploy the track segmentation model to a real-world application environment and execute an adaptive online learning process, wherein the adaptive online learning process includes: continuously collecting newly arrived track data, constructing a dynamic database, and periodically fine-tuning the track segmentation model online using data from the dynamic database and a preset hybrid loss function.
[0007] According to some embodiments of the present invention, in step S2, the mask reconstruction task specifically includes: randomly masking the state information of some track points in the unlabeled dataset, using the model to predict the masked state information, and calculating the reconstruction error through a first loss function.
[0008] In some embodiments of the present invention, step S2 specifically includes: performing different data augmentations on the same track data to generate positive sample pairs; generating negative sample pairs from data of different tracks; extracting features using the model and calculating similarity; and using a second loss function to bring the features of positive sample pairs closer together and pull the features of negative sample pairs further apart.
[0009] According to some embodiments of the present invention, in step S2, the total loss function of the self-supervised pre-training is a weighted sum of the first loss function of the mask reconstruction task and the second loss function of the contrastive learning task.
[0010] In some embodiments of the present invention, in step S3, the loss function used for supervised fine-tuning is a focus loss function.
[0011] According to some embodiments of the present invention, in step S3, the encoder part of the pre-trained model is frozen in layers, and only the later layers of the encoder and the newly added segmentation decision head are fine-tuned.
[0012] In some embodiments of the present invention, step S4, specifically includes: storing only manually labeled new data or new data with model prediction confidence higher than a set threshold into the database; the database has a maximum capacity, and when the data volume reaches the maximum capacity, the earliest stored data is removed in chronological order.
[0013] According to some embodiments of the present invention, in step S4, the preset hybrid loss function is a weighted sum of the first loss function of the mask reconstruction task and the loss function used in the supervised fine-tuning.
[0014] According to an embodiment of the present invention, a trajectory segmentation system based on spatiotemporal awareness and adaptive learning includes: The preprocessing module is used to preprocess the raw track data to obtain unlabeled and labeled datasets; The pre-training module is used to build a Transformer-based encoder-decoder model. Using the unlabeled dataset, the model is pre-trained in a self-supervised manner through mask reconstruction and contrastive learning tasks to obtain a pre-trained model. The fine-tuning module is used to add a segmentation decision head to the pre-trained model and perform supervised fine-tuning using the labeled dataset to obtain the track segmentation model. The deployment and online learning module is used to deploy the track segmentation model and perform an adaptive online learning process, which includes: continuously collecting newly arriving track data, building a dynamic database, and periodically fine-tuning the track segmentation model online using data in the dynamic database and a preset hybrid loss function.
[0015] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the trajectory segmentation method based on spatiotemporal awareness and adaptive learning as described above.
[0016] The present invention has the following beneficial effects: First, this invention employs a "pre-training-fine-tuning-continuous learning" approach, fully utilizing unlabeled data for model pre-training. Only a small amount of data needs to be labeled, significantly shortening the time required for data preparation and effectively reducing the time cost of training the track segmentation model. Furthermore, fine-tuning the model using a small amount of labeled data allows for rapid adaptation to specific track segmentation tasks. Compared to traditional methods, this approach offers significantly higher data efficiency and enables rapid deployment of the model to real-world application environments.
[0017] Secondly, this invention learns universal flight path motion patterns through pre-training, rather than superficial features of a specific dataset, thus enabling better generalization to new, unseen scenarios and targets. Simultaneously, continuous online learning ensures the model can continuously adapt to changes in data distribution, maintaining excellent performance over the long term. Compared to traditional models, this model has stronger generalization ability and stability, while also compensating for the time-dependent stability degradation of traditional neural network models, providing more reliable technical support for air traffic management and military defense.
[0018] Third, this invention also uses Transformer instead of traditional recurrent neural network, and its attention weights can be visualized to analyze which key points and timestamps the model focuses on when making segmentation decisions. This can provide a basis for algorithm decision-making, increase the credibility and interpretability of segmentation results, make it easier for engineers to understand and trust the model output, and at the same time provide assistance for the research on trajectory prediction problems. Attached Figure Description
[0019] Figure 1 This is a flowchart of a trajectory segmentation method based on spatiotemporal awareness and adaptive learning according to an embodiment of the present invention; Figure 2 This is a schematic diagram of track outlier values according to an embodiment of the present invention; Figure 3 This is a flowchart of the self-supervised pre-training process according to an embodiment of the present invention. Detailed Implementation
[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments.
[0021] The steps described in the specification and the flowcharts in the accompanying drawings of this invention are not necessarily strictly executed according to the step numbers; the execution order of the steps can be changed. Furthermore, certain steps can be omitted, multiple steps can be combined into one step, and / or one step can be broken down into multiple steps.
[0022] This invention proposes a track segmentation method based on spatiotemporal awareness and adaptive learning to address the problems of high time cost, poor model generalization ability, and poor interpretability in solving track segmentation problems using traditional deep learning methods. Specifically, in the data preparation stage, a multi-dimensional spatiotemporal awareness module is constructed through mask reconstruction and contrastive learning, enabling the model to be pre-trained through self-supervised learning, thereby reducing the time required for large-scale labeled datasets. In the model training stage, a large-scale unlabeled dataset is first used for pre-training, making it easier for the model to learn high-quality track features with strong generalization performance. After pre-training, a small-scale labeled dataset is used for model fine-tuning, while an attention mechanism and a segmentation decision head are introduced to adapt its general track knowledge to specific segmentation tasks, further enhancing its segmentation performance and generalization ability. After training, the model is deployed and continuously fine-tuned using unlabeled data with the addition of a dynamic selection strategy, enabling the model to adapt to environmental changes and new patterns. In addition, since the attention mechanism can be visualized, it can show which track points and timestamps the model pays attention to when making decisions, enhancing the model's interpretability.
[0023] This invention can be used in ground analysis systems to segment the trajectory of aerial targets for air traffic control, and can also help with the identification of aerial target intentions and trajectory prediction.
[0024] The present invention will now be further described with reference to the accompanying drawings.
[0025] See attached document Figure 1 This invention proposes a trajectory segmentation method based on spatiotemporal awareness and adaptive learning, comprising the following steps: Step 1: Prepare raw data and perform data preprocessing to provide high-quality, uniformly formatted input data for subsequent training. The raw data consists of flight track data from a historical database. Each track is a time series, and each time point contains the target's status information, such as timestamp, latitude and longitude, altitude, and speed. 80% of the track data is selected as unlabeled data, while the remaining 20% can be labeled by engineers. Each track point is assigned a segmentation label, such as turning, takeoff, or hovering, to indicate which track segment it belongs to.
[0026] After the track data is prepared, it needs to be preprocessed and enhanced to improve its stability. First, the Local Outlier Factor (LOF) method is used to remove outliers from the track data. In the LOF method, the LOF of each data point is calculated to determine whether that data point is an outlier. The LOF value is derived from the Local Reachability Density (LRD) of each point. The calculation method for LRD is as follows: ; in, Point The Distance neighborhood; Represents the midpoint of space Time The Reachable distance.
[0027] The formula for calculating the LOF value at this point is as follows: ; The LOF value is the ratio of the LRD value of the measurement point to the average LRD value of all its neighboring measurement points.
[0028] The LOF algorithm is suitable for processing sequential data. Its sensitivity can be adjusted by changing the number of neighboring points and the distance threshold, and it exhibits good stability in identifying outliers. This invention uses the LOF algorithm to identify outliers in a dataset. If the LOF value of a point is greater than 1, it is identified as an outlier and deleted. Outliers in flight tracks are as follows: Figure 2 As shown.
[0029] Then, cubic spline interpolation was used to fill in the missing values. Let the function be a quadratic continuously differentiable function. interval If divided into n consecutive intervals, then there are a total of There are points, among which , The interval distribution is shown below: ; Use piecewise polynomials within each interval As shown below: ; Will It is believed to be A cubic spline interpolation function only needs to satisfy the following condition: The polynomials in each interval are either of degree no higher or zero, as shown below: ; When filling in dotted data, the inserted data should be consistent with the original data, as shown below: ; in, =1, 2, ..., n+1.
[0030] It is a twice continuously differentiable function, as shown below: ; in, =1, 2, ..., n-1.
[0031] If the calculated fill value is negative, then the minimum value of the function is taken as the fill value, as shown below: ; Finally, the track data is normalized. This invention uses the max-min normalization method to normalize the track data, and the calculation method is as follows: ; Where X represents the data after data normalization; x represents the data before data normalization. The minimum value of the corresponding track feature data; This represents the maximum value of the corresponding track feature data.
[0032] After preprocessing, the unlabeled dataset is shown below: ; in, This is an unlabeled track dataset; These are the dots in the flight path.
[0033] After preprocessing, the labeled dataset is as follows: ; Each dot contains characteristic information about the aircraft type, including time, longitude, latitude, altitude, speed, and heading angle, as shown in the following formula: ; in, For the time of the dot; Longitude of the dot; The latitude of the dot; The height of the dot; For the speed of the dot; The heading angle of the dot; The pitch angle of the dot is denoted by .
[0034] Step 2: Self-supervised pre-training The flowchart of self-supervised pre-training is as follows: Figure 3 As shown.
[0035] First, the model architecture is designed. This invention employs a Transformer-based encoder-decoder architecture, equipped with a contrastive learning projector. The backbone encoder is a standard Transformer Encoder with added multi-layer self-attention mechanisms to map each trackpoint to a high-dimensional space. The pre-training task decoder uses a lightweight Transformer Decoder for pre-training task one (mask reconstruction). It reconstructs the coordinates of masked trackpoints by inputting the decoder's output and a special token sequence. The contrastive learning projector is a multi-layer dry preparation for pre-training task two (contrastive learning). It takes track features as input and outputs a low-dimensional, normalized representation vector for calculating track feature similarity. After the model architecture design is complete, pre-training tasks one and two can begin.
[0036] Pre-training Task 1: Mask Reconstruction. Retrieve preprocessed, unlabeled track data. Randomly select 15% of the waypoints and replace them using the following strategy: 80% probability of replacing with a trainable special token sequence, 10% probability of replacing with random points, and 10% probability of leaving it unchanged. This yields the corrupted input sequence. .Will Input the backbone encoder to obtain contextual features ,Will The decoder is pre-trained with a special token sequence to predict the masked original coordinates. In pre-training task one, the loss function uses a smooth L1 loss function that is less sensitive to outliers, and its calculation formula is: ; Pre-training Task 2: Comparative Learning. First, construct the input sequence, and then perform comparative learning on the same batch of tracks. Two different data augmentation operations were performed, one adding random white noise and the other adding Gaussian noise, to obtain the results. and ,in The track data after adding Gaussian noise. This refers to the track data after adding random white noise. Then, [then...] and The output sequences are obtained by inputting the core encoder separately. After performing global average pooling on the features of this sequence, it is input into the contrastive learning projection head to obtain the output vector. and In the same batch, and It is a positive sample pair, and and Other vectors in the same batch and (k≠i) are all negative sample pairs. In pre-training task two, the loss function used is the normalized temperature-scaled cross-entropy loss function, which is calculated as follows: ; in, Calculate the cosine similarity; This is a temperature parameter used to control the sharpness of the distribution; This is an indicator function; it is 0 when k=i, and 1 otherwise.
[0037] Finally, multi-task joint training is performed. The weighted sum of the two loss functions is used as the overall pre-training objective. The formula for the total loss is shown below: ; Among them, hyperparameters To balance two pre-training tasks, it is typically set to... Starting from this point, the backbone encoder weights are updated, and then fine-tuned based on the validation set.
[0038] Step 3: Supervised fine-tuning; The general track feature extractor obtained in step two is applied to the specific task of multi-target track segmentation, enabling it to assign a correct target label to each point in the sequence.
[0039] First, a "task head" for segmentation is constructed based on the pre-trained model. Weights from the backbone encoder are loaded, and then the first 10% of the backbone encoder's layers are frozen to prevent the loss of learned general track features after training on a small labeled dataset. Next, a multi-head attention mechanism is added to the backbone encoder, ensuring that each point's features contain contextual information about the global sequence, explicitly modeling the spatiotemporal relationships between points. Finally, a multilayer perceptron is used to map the high-dimensional features to segmentation labels, obtaining the final segmentation result.
[0040] In track segmentation tasks, the imbalance in the number of points within each track segment often leads to a bias towards the majority class when using standard cross-entropy loss. Therefore, a focus loss function is used at this stage, calculated as follows: ; in, Used to balance class weights, allowing for larger values to be set for minority classes; This is a focus parameter, typically set to 2; It is the model's predicted probability for the sample.
[0041] Step 4: Model Deployment and Inference. The ONNX tool is used to accelerate and optimize the model, improving its real-time segmentation capabilities. After the model outputs track labels for each point, the point sequences are grouped according to the labels to form different track segments. If the number of points in a track segment is less than 10, the track segment is merged into the next track segment; if there is no next track segment, the track segment is merged into the previous track segment.
[0042] Step 5: Adaptive Online Learning. After the model is deployed to the real environment, it continuously and lightweightly learns from newly arriving data to address the problem of data distribution drift and adapt to new motion patterns not seen in the training set, thereby maintaining and improving model performance and stability.
[0043] First, a dynamic database is constructed. Because new data may be too large, contain noise, and contain errors, a database is needed to store valuable new samples. This database only stores manually labeled data and data with a model prediction confidence level higher than 90%. The number of data points for each motion mode in the database is statistically analyzed and balanced to avoid duplicate storage of simple samples. Simultaneously, a maximum database capacity is set; when the database reaches its maximum capacity, the oldest data samples are removed to improve data timeliness. After the dynamic database is constructed, adaptive model updates can begin.
[0044] Adaptive model updates. This invention periodically fine-tunes the model using data from the database, i.e., using a very small learning rate and performing a few training rounds. Regarding the design of the loss function, to simultaneously preserve existing knowledge and learned knowledge, this invention combines the mask reconstruction loss function from step two and the focus loss function from step three, with the calculation formula shown below: ; in, This is the total loss function; Output the loss function value for sample data with high confidence for the model; The loss function value is for sample data with manually labeled data; , , For hyperparameters, It is usually 1. .
[0045] In summary, this invention constructs a complete closed-loop learning paradigm of "pre-training-fine-tuning-continuous learning," and its key innovations are as follows: First, this invention designs a dual self-supervised training task. Through mask reconstruction, the model learns the local structure and motion patterns of the track; through contrastive learning, the model understands the global semantic similarity of different rows, enabling it to learn rich track knowledge from unlabeled data. Employing a "self-supervised pre-training - small-sample fine-tuning" training mode, the powerful general-purpose model obtained through pre-training can be quickly fine-tuned using a small amount of labeled data, accurately adapting to specific track segmentation tasks. Compared to traditional methods, this allows for faster model training, saving time and solving the problem of excessively high time costs in the data preparation stage of traditional methods.
[0046] Secondly, this invention designs a spatiotemporal context-aware hybrid neural network, capable of accurately modeling complex scenes. By using a Transformer instead of a traditional recurrent neural network, the model can directly capture long-range dependencies and global contextual relationships between waypoints. Compared to traditional recognition methods, this further ensures accurate segmentation of aerial target tracks in practical applications, especially in complex environments, enhancing the system's real-time performance and practicality, and enabling the system to better handle track segmentation tasks in complex environments.
[0047] Third, this invention also establishes a continuous adaptive online learning system, enabling the model to learn throughout its life. By constructing a dynamic database and establishing a hybrid loss function for online fine-tuning, the long-term stability of the model is further improved. Compared to the performance degradation over time characteristic of traditional deep learning models, this invention consolidates its advantages in addressing the shortcomings of traditional methods, providing more reliable technical support for air traffic management and military defense.
[0048] The present invention has the following beneficial effects: First, this invention employs a "pre-training-fine-tuning-continuous learning" approach, fully utilizing unlabeled data for model pre-training. Only a small amount of data needs to be labeled, significantly shortening the time required for data preparation and effectively reducing the time cost of training the track segmentation model. Furthermore, fine-tuning the model using a small amount of labeled data allows for rapid adaptation to specific track segmentation tasks. Compared to traditional methods, this approach offers significantly higher data efficiency and enables rapid deployment of the model to real-world application environments.
[0049] Secondly, this invention learns universal flight path motion patterns through pre-training, rather than superficial features of a specific dataset, thus enabling better generalization to new, unseen scenarios and targets. Simultaneously, continuous online learning ensures the model can continuously adapt to changes in data distribution, maintaining excellent performance over the long term. Compared to traditional models, this model has stronger generalization ability and stability, while also compensating for the time-dependent stability degradation of traditional neural network models, providing more reliable technical support for air traffic management and military defense.
[0050] Third, this invention also uses Transformer instead of traditional recurrent neural network, and its attention weights can be visualized to analyze which key points and timestamps the model focuses on when making segmentation decisions. This can provide a basis for algorithm decision-making, increase the credibility and interpretability of segmentation results, make it easier for engineers to understand and trust the model output, and at the same time provide assistance for the research on trajectory prediction problems.
[0051] Through the description of specific embodiments, a more in-depth and specific understanding should be gained of the technical means and effects adopted by the present invention to achieve the intended purpose. However, the accompanying drawings are only provided for reference and illustration and are not intended to limit the present invention.
Claims
1. A trajectory segmentation method based on spatiotemporal awareness and adaptive learning, characterized in that, include: S1, preprocess the original track data to obtain unlabeled dataset and labeled dataset; S2, Construct a Transformer-based encoder-decoder model, and use the unlabeled dataset to perform self-supervised pre-training on the model through mask reconstruction and contrastive learning tasks to obtain a pre-trained model; S3. Based on the pre-trained model, add a segmentation decision head and perform supervised fine-tuning using the labeled dataset to obtain the track segmentation model. S4, Deploy the track segmentation model to a real-world application environment and execute an adaptive online learning process, wherein the adaptive online learning process includes: continuously collecting newly arrived track data, constructing a dynamic database, and periodically fine-tuning the track segmentation model online using data from the dynamic database and a preset hybrid loss function.
2. The trajectory segmentation method based on spatiotemporal awareness and adaptive learning according to claim 1, characterized in that, In step S2, the mask reconstruction task specifically includes: randomly masking the state information of some track points in the unlabeled dataset, using the model to predict the masked state information, and calculating the reconstruction error through a first loss function.
3. The trajectory segmentation method based on spatiotemporal awareness and adaptive learning according to claim 1 or 2, characterized in that, In step S2, the contrastive learning task specifically includes: performing different data augmentations on the same track data to generate positive sample pairs; generating negative sample pairs from data of different tracks; extracting features using the model and calculating similarity; and using a second loss function to bring the features of positive sample pairs closer together and pull the features of negative sample pairs further apart.
4. The trajectory segmentation method based on spatiotemporal awareness and adaptive learning according to claim 1, characterized in that, In step S2, the total loss function of the self-supervised pre-training is the weighted sum of the first loss function of the mask reconstruction task and the second loss function of the contrastive learning task.
5. The trajectory segmentation method based on spatiotemporal awareness and adaptive learning according to claim 1, characterized in that, In step S3, the loss function used in the supervised fine-tuning is the focus loss function.
6. The trajectory segmentation method based on spatiotemporal awareness and adaptive learning according to claim 1, characterized in that, In step S3, the encoder part of the pre-trained model is frozen in layers, and only the later layers of the encoder and the newly added segmentation decision head are fine-tuned.
7. The trajectory segmentation method based on spatiotemporal awareness and adaptive learning according to claim 1, characterized in that, In step S4, the construction of the dynamic database specifically includes: storing only manually labeled new data or new data with model prediction confidence higher than a set threshold into the database; the database has a maximum capacity, and when the data volume reaches the maximum capacity, the earliest stored data is removed in chronological order.
8. The trajectory segmentation method based on spatiotemporal awareness and adaptive learning according to claim 1, characterized in that, In step S4, the preset hybrid loss function is the weighted sum of the first loss function of the mask reconstruction task and the loss function used in the supervised fine-tuning.
9. A trajectory segmentation system based on spatiotemporal awareness and adaptive learning, characterized in that, include: The preprocessing module is used to preprocess the raw track data to obtain unlabeled and labeled datasets; The pre-training module is used to build a Transformer-based encoder-decoder model. Using the unlabeled dataset, the model is pre-trained in a self-supervised manner through mask reconstruction and contrastive learning tasks to obtain a pre-trained model. The fine-tuning module is used to add a segmentation decision head to the pre-trained model and perform supervised fine-tuning using the labeled dataset to obtain the track segmentation model. The deployment and online learning module is used to deploy the track segmentation model and perform an adaptive online learning process, which includes: continuously collecting newly arriving track data, building a dynamic database, and periodically fine-tuning the track segmentation model online using data in the dynamic database and a preset hybrid loss function.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the track segmentation method based on spatiotemporal awareness and adaptive learning as described in any one of claims 1 to 8.