Rail transit obstacle recognition method based on small sample continuous learning model

Abstract

The present invention provides a rail transit obstacle recognition method based on a small sample continuous learning model, which belongs to the field of rail transit technology. The method includes: acquiring a target image, preprocessing the target image, and obtaining a preprocessed target image; extracting a track limit area image from the preprocessed target image; inputting the track limit area image into a pre-constructed small sample continuous learning model and / or an updated small sample continuous learning model, and obtaining an obstacle recognition result corresponding to the track limit area image output by the small sample continuous learning model and / or the updated small sample continuous learning model. The rail transit obstacle recognition method based on a small sample continuous learning model provided by the present invention can effectively identify new track obstacles and is suitable for large-scale and complex application scenarios.

Description

Technical Field

[0001] The present invention relates to the field of rail transit technology, and in particular to a rail transit obstacle recognition method based on a small sample continuous learning model. Background Art

[0002] Rail transit safety is closely related to urban development. Track obstacles not only hinder the normal operation of the transportation system, but may also cause various accidents, posing a threat to passengers and traffic safety. Therefore, it is very important to detect and classify rail transit obstacles. At present, traditional target detection models are usually used to identify rail transit obstacles. However, traditional methods perform poorly when dealing with new types of obstacles and require a large amount of sample data for training. In practical applications, due to the limited samples of specific types of obstacles, traditional methods have difficulty in identifying new types of obstacles, especially when facing large-scale and complex rail transit systems, the generalization performance of the model is challenged. Summary of the invention

[0003] The present invention provides a rail transit obstacle recognition method based on a small sample continuous learning model, which is used to solve the defects of the prior art that a traditional target detection model is used to recognize rail transit obstacles, it is difficult to recognize new obstacles, and it is not suitable for large-scale and complex application scenarios.

[0004] The present invention provides a rail transit obstacle recognition method based on a small sample continuous learning model, comprising:

[0005] Acquire a target image, and preprocess the target image to obtain a preprocessed target image;

[0006] Extracting a track limit area image from the preprocessed target image;

[0007] Inputting the track limit area image into a pre-built small sample continuous learning model and / or an updated small sample continuous learning model, and obtaining an obstacle recognition result corresponding to the track limit area image output by the small sample continuous learning model and / or the updated small sample continuous learning model;

[0008] The small sample continuous learning model is obtained by training based on old sample images of the track limit area and labels of obstacle recognition results corresponding to the old sample images of the track limit area;

[0009] Among them, the updated small sample continuous learning model is obtained by continuously training the small sample continuous learning model based on the new sample images of the track limit area and the old sample images of the track limit area, as well as the obstacle recognition result labels corresponding to the new sample images of the track limit area and the obstacle recognition result labels corresponding to the old sample images of the track limit area, combined with the long short-term memory mechanism.

[0010] In some embodiments, before extracting the track limit area image from the pre-processed target image, the method further includes:

[0011] Acquire a reference image sequence, and preprocess the reference image sequence to obtain a preprocessed reference image sequence;

[0012] Extracting a track limit area reference image from the preprocessed reference image sequence;

[0013] Correspondingly, extracting the track limit area image from the preprocessed target image includes:

[0014] The preprocessed target image is spatially aligned with the track limit area reference image, and the track limit area image is extracted from the preprocessed target image.

[0015] In some embodiments, before inputting the track limit area image into the small sample continuous learning model, the method further includes:

[0016] Acquire target radar data, and preprocess the target radar data to obtain preprocessed target radar data;

[0017] Based on the preprocessed target radar data, the track limit area image is classified using an open world algorithm to obtain a classification result, where the classification result includes known categories and / or unknown categories.

[0018] In some embodiments, the method includes: when it is determined that the classification result includes a known category, inputting the track limit area image into the pre-built small sample continuous learning model to obtain an obstacle recognition result corresponding to the track limit area image output by the small sample continuous learning model.

[0019] In some embodiments, the method includes: when it is determined that the classification result includes an unknown category, inputting the track limit area image into the updated small sample continuous learning model to obtain an obstacle recognition result corresponding to the track limit area image output by the updated small sample continuous learning model.

[0020] In some embodiments, the continuously training the small sample continuous learning model includes:

[0021] Initializing parameters of the small sample continuous learning model;

[0022] The old sample images of the track limit area are stored in a memory storage space, and the old sample images of the track limit area are randomly extracted from the memory storage space to obtain a memory playback sample image set;

[0023] The memory playback sample image set is mixed with the track limit area new sample image to obtain a training sample image set, and a training sample image is randomly extracted from the training sample image set;

[0024] In the case where the extracted training sample image is the new sample image of the track limit area, based on the new sample image of the track limit area, performing stochastic gradient descent on the parameters of the small sample continuous learning model, and updating the current projection operator;

[0025] In the case where the extracted training sample image is the memory playback sample image, performing stochastic gradient descent on the parameters of the small sample continuous learning model based on the memory playback sample image;

[0026] Updating the parameters of the small sample continuous learning model to obtain an updated small sample continuous learning model;

[0027] The new sample image of the track limit area is stored in the memory storage space.

[0028] In some embodiments, the continuously training the small sample continuous learning model includes:

[0029] Extracting unknown category candidate regions from the new sample image of the track limit area, and determining unknown category target regions from the unknown category candidate regions;

[0030] Extracting features from the target area to obtain a new feature vector;

[0031] Randomly selecting a memory playback sample image from the old sample image of the track limit area, performing feature extraction on the memory playback sample image, and obtaining a reconstructed feature vector;

[0032] Fusing the new feature vector and the reconstructed feature vector to obtain a fused feature vector;

[0033] Pooling the fused feature vector to obtain a pooled feature vector;

[0034] Based on the pooled feature vector, predicting the obstacle recognition result corresponding to the new sample image of the track limit area to obtain a prediction result;

[0035] Based on the prediction result and the obstacle recognition result label corresponding to the new sample image of the track limit area, a loss function value is calculated, and based on the loss function value, the parameters of the small sample continuous learning model are updated.

[0036] In some embodiments, storing the new sample image of the track limit area in the memory storage space includes:

[0037] Extracting features of the new sample image of the track limit area to obtain a feature vector;

[0038] The feature vector is compressed to obtain a compressed feature vector, and the compressed feature vector is stored in the memory storage space.

[0039] In some embodiments, the expression for performing stochastic gradient descent on the parameters of the small sample continuous learning model based on the new sample image of the track limit area is as follows:

[0040] θ i =SGD(x i1 ,y i1 ,θ i-1 ,sβ)

[0041] Among them, θ i represents the parameter of the i-th gradient of the small sample continuous learning model, θ i-1 represents the parameter of the i-1th gradient of the small sample continuous learning model, s represents the attention coefficient, β represents the inner loop learning rate, and x i1 represents the new sample image of the track limit area, y i1 Indicates the obstacle recognition result label corresponding to the new sample image of the track limit area;

[0042] The expression for performing stochastic gradient descent on the parameters of the small sample continuous learning model based on the memory playback sample image is as follows:

[0043] θ i =SGD(x i2 ,y i2 ,θ i-1 ,β)

[0044] Among them, x i2 represents the memory playback sample image, y i2 Indicates the obstacle recognition result label corresponding to the memory playback sample image.

[0045] In some embodiments, the calculation formula for updating the parameters of the small sample continuous learning model is as follows:

[0046] θ=θ0+θ*(θi -θ0)P m

[0047] Wherein, θ represents the updated parameters of the small sample continuous learning model, θ0 represents the initial parameters of the small sample continuous learning model, and θ i represents the parameter of the first gradient of the small sample continuous learning model, γ represents the outer loop learning rate, P m Represents the updated projection operator.

[0048] The present invention provides a rail transit obstacle recognition method based on a small sample continuous learning model, which obtains a target image, preprocesses the target image, and obtains a preprocessed target image; extracts a track limit area image from the preprocessed target image; and based on the track limit area image, obtains an obstacle recognition result corresponding to the track limit area image using a pre-constructed small sample continuous learning model and / or an updated small sample continuous learning model. The method can effectively identify new obstacles and is suitable for large-scale and complex application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 This is one of the flow charts of the rail transit obstacle identification method based on the small sample continuous learning model provided by the embodiment of the present invention;

[0051] Figure 2 This is the second flow chart of the rail transit obstacle identification method based on the small sample continuous learning model provided by the embodiment of the present invention;

[0052] Figure 3 It is one of the flow charts of continuous training of a small sample continuous learning model provided by the real-time example of the present invention;

[0053] Figure 4 This is the second flowchart of training a small sample continuous learning model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] In recent years, traffic obstacle detection based on machine learning and deep learning has been widely used in the field of unmanned driving. However, for the current detection algorithm, the main difficulty lies in data imbalance and incompleteness. Under normal circumstances, the stock of obstructed track data is much lower than that of normal track data. This means that obstacle data occupies a small proportion in the data set. At present, the application of rail transit obstacle detection in actual scenarios is still relatively limited. On the one hand, the infrastructure is not yet perfect, but another part of the reason is that the versatility and application ability of related technologies are poor. The main defects of related technologies include the following aspects: 1) As a deep learning model for traffic obstacle detection, it is difficult to cope with a variety of possible unseen obstacles; 2) The collection of data sets related to rail transit obstacles is difficult, the data samples are small, and the model is difficult to effectively use the data.

[0056] In addition, due to the high speed of common rail trains, the short processing time of the detection system, and the complex and changeable scenes, it is difficult for machine learning-based methods to achieve satisfactory results in terms of accuracy and speed. At the same time, since most of the current rail obstacle detection methods are based on supervised learning methods, they cannot identify targets that do not exist in the pre-training data set, so it is difficult to apply them when the target is uncertain.

[0057] To this end, an embodiment of the present invention provides a rail transit obstacle recognition method based on a small sample continuous learning model, by acquiring a target image, preprocessing the target image, and obtaining a preprocessed target image; extracting a track limit area image from the preprocessed target image; based on the track limit area image, using a pre-constructed small sample continuous learning model and / or an updated small sample continuous learning model, obtaining an obstacle recognition result corresponding to the track limit area image. The present invention can effectively identify new types of obstacles and is suitable for large-scale and complex application scenarios.

[0058] Figure 1 One of the flow charts of the rail transit obstacle identification method based on the small sample continuous learning model provided by the embodiment of the present invention. Figure 1 As shown, a rail transit obstacle identification method based on a small sample continuous learning model is provided, comprising the following steps: step 110, step 120, and step 130. The method flow steps are only a possible implementation of the present invention.

[0059] Step 110: Acquire a target image, and preprocess the target image to obtain a preprocessed target image.

[0060] Optionally, the target image can be acquired by a front camera mounted on the vehicle.

[0061] Optionally, the target image is preprocessed by denoising, graying, etc. to obtain a preprocessed target image.

[0062] Step 120: extracting the track limit area image from the preprocessed target image.

[0063] The track limit area image refers to the image of the track boundary range, indicating the area where the train can travel.

[0064] In some embodiments, before extracting the track limit area image from the pre-processed target image, the method further includes:

[0065] Acquire a reference image sequence, preprocess the reference image sequence, and obtain a preprocessed reference image sequence;

[0066] Extracting a track limit area reference image from the preprocessed reference image sequence;

[0067] Correspondingly, the track limit area image is extracted from the preprocessed target image, including:

[0068] The preprocessed target image is spatially aligned with the track limit area reference image, and the track limit area image is extracted from the preprocessed target image.

[0069] It should be noted that by preprocessing the reference image sequence and extracting the track limit area reference image from the preprocessed reference image sequence, relatively stable reference points can be obtained. These reference points can be used as a benchmark for spatial alignment of the target image and the reference image, thereby ensuring that the extracted track limit area image is consistent with the actual track position.

[0070] It can be understood that by spatially aligning the preprocessed target image with the track limit area reference image and extracting the track limit area image from the preprocessed target image, the quality and accuracy of the track limit area image can be improved, thereby providing a more reliable data basis for subsequent rail transit obstacle identification.

[0071] Step 130: Input the track limit area image into a pre-built small sample continuous learning model and / or an updated small sample continuous learning model to obtain an obstacle recognition result corresponding to the track limit area image output by the small sample continuous learning model and / or the updated small sample continuous learning model.

[0072] Among them, the small sample continuous learning model is trained based on the old sample images of the track limit area and the obstacle recognition result labels corresponding to the old sample images of the track limit area.

[0073] Among them, the updated small sample continuous learning model is based on the new sample images of the track limit area and the old sample images of the track limit area, as well as the obstacle recognition result labels corresponding to the new sample images of the track limit area and the obstacle recognition result labels corresponding to the old sample images of the track limit area, combined with the long short-term memory mechanism, to continuously train the small sample continuous learning model.

[0074] It can be understood that using a small sample continuous learning model and / or an updated small sample continuous learning model to identify obstacles within the track limit area can improve the accuracy and efficiency of identification and help ensure the safe operation of the rail transit system.

[0075] In some embodiments, before inputting the track limit area image into the small sample continuous learning model, the method further includes:

[0076] Acquire target radar data, preprocess the target radar data, and obtain preprocessed target radar data;

[0077] Based on the preprocessed target radar data, the track limit area image is classified using an open world algorithm to obtain a classification result, which includes known categories and / or unknown categories.

[0078] Among them, the target radar data includes information such as the position, speed, size and shape of the target object. By obtaining the target radar data, more comprehensive and multi-angle track limit area information can be obtained, which helps to improve obstacle recognition capabilities, especially in severe weather or lighting conditions.

[0079] Optionally, the target radar data is preprocessed into grayscale to obtain a grayscale image of the target radar data.

[0080] It should be noted that the open-world algorithm can recognize known and unknown categories, which helps to identify new types of obstacles or abnormal situations, so that timely measures can be taken to avoid accidents.

[0081] It can be understood that combining the preprocessed target radar data with the track limit area image and classifying the track limit area image using the open world algorithm can improve the system's multimodal perception, robustness and early warning capabilities, thereby better ensuring the safe operation of the rail transit system.

[0082] In some embodiments, the method includes: when it is determined that the classification result includes a known category, inputting the track limit area image into a pre-built small sample continuous learning model to obtain an obstacle recognition result corresponding to the track limit area image output by the small sample continuous learning model.

[0083] It can be understood that, when it is determined that the track limit area image includes known categories, directly using the pre-built small sample continuous learning model can quickly identify obstacles and improve the efficiency of obstacle recognition.

[0084] In some embodiments, the method includes: when it is determined that the classification result includes an unknown category, inputting the track limit area image into an updated small sample continuous learning model to obtain an obstacle recognition result corresponding to the track limit area image output by the updated small sample continuous learning model.

[0085] It can be understood that when it is determined that the track limit area image includes unknown categories, the updated small sample continuous learning model can be used to identify newly emerging obstacle types and can flexibly respond to the challenges of diverse and frequently changing obstacle types in the rail transit environment.

[0086] Figure 2 The second flow chart of the rail transit obstacle identification method based on the small sample continuous learning model provided by the embodiment of the present invention. Figure 2 As shown, a rail transit obstacle recognition method based on a small sample continuous learning model is provided, including a preprocessing stage, an obstacle recognition stage and an auxiliary processing stage.

[0087] In the preprocessing stage, the reference video image sequence is preprocessed to extract the track boundary area, obtain the track reference information and old sample images and store them in different databases.

[0088] In the obstacle recognition stage, radar data and video images are acquired, radar data are preprocessed to obtain preprocessed radar data, video images are preprocessed to obtain preprocessed target images, the preprocessed target images and reference images are spatially aligned, track limit area images are extracted from the preprocessed target images, the preprocessed radar data and track limit area images are fused, the track limit area images are classified in combination with the open world algorithm to obtain classification results, when the classification results include known classes, multimodal target detection is performed on the track limit area images, and the detection results are sent to the train control system, and when the classification results include unknown classes, multimodal target detection is performed on the track limit area images in combination with new sample images and old sample images, and the detection results are sent to the train control system.

[0089] In the auxiliary processing stage, when the classification results include unknown classes, artificial intelligence methods are used to automatically generate labels for new sample images, and alignment is performed for manual correction to obtain labeled new sample images.

[0090] In an embodiment of the present invention, a target image is acquired and preprocessed to obtain a preprocessed target image; a track limit area image is extracted from the preprocessed target image; based on the track limit area image, a pre-constructed small sample continuous learning model and / or an updated small sample continuous learning model is used to obtain an obstacle recognition result corresponding to the track limit area image, which can effectively identify new obstacles and is suitable for large-scale and complex application scenarios.

[0091] It should be noted that each implementation of the present invention can be freely combined, the order can be changed, or it can be executed separately, and does not need to rely on or depend on a fixed execution order.

[0092] Figure 3 One of the flow charts of continuous training of a small sample continuous learning model provided by the present invention is as follows: Figure 3 As shown, in some embodiments, the small sample continuous learning model is continuously trained, including steps 310 to 370.

[0093] Step 310: Initialize the parameters of the small sample continuous learning model.

[0094] Step 320: Store old sample images of the track limit area in a memory storage space, and randomly extract old sample images of the track limit area from the memory storage space to obtain a memory playback sample image set.

[0095] Step 330: Mix the memory playback sample image set with the new sample image of the track limit area to obtain a training sample image set, and randomly extract a training sample image from the training sample image set.

[0096] It can be understood that by mixing old sample images with new sample images to form a training sample image set, the diversity of training data can be increased, which helps the model to better learn and generalize to different situations and scenarios, and improves the robustness and generalization of the model.

[0097] Step 340: When the extracted training sample image is a new sample image of the track limit area, based on the new sample image of the track limit area, perform stochastic gradient descent on the parameters of the small sample continuous learning model and update the current projection operator.

[0098] Step 350: When the extracted training sample image is a memory playback sample image, perform stochastic gradient descent on the parameters of the small sample continuous learning model based on the memory playback sample image.

[0099] It can be understood that personalized parameter updates can be achieved by selecting different parameter update strategies according to the source of the training sample images; for new sample images in the track limit area, stochastic gradient descent is used to update the current projection operator, so that the model can better adapt to new data; for memory playback sample images, stochastic gradient descent is used to update the parameters of the small sample continuous learning model, so as to maintain the adaptability to old tasks.

[0100] In some embodiments, based on the new sample image of the track boundary area, the expression for stochastic gradient descent of the parameters of the small sample continuous learning model is as follows:

[0101] θ i =SGD(x i1 ,y i1 ,θ i-1 ,sβ)

[0102] Among them, θ i represents the parameter of the i-th gradient of the small sample continuous learning model, θ i-1 represents the parameter of the i-1th gradient of the small sample continuous learning model, s represents the attention coefficient, β represents the inner loop learning rate, and x i1 Represents the new sample image of the track limit area, y i1 Indicates the obstacle recognition result label corresponding to the new sample image of the track limit area;

[0103] Based on the memory playback sample image, the expression of stochastic gradient descent for the parameters of the small sample continuous learning model is as follows:

[0104] θ i =SGD(x i2 ,y i2 ,θ i-1 ,β)

[0105] Among them, x i2 Represents the memory playback sample image, y i2 Indicates the obstacle recognition result label corresponding to the memory playback sample image.

[0106] It should be noted that the attention coefficient s based on the number of replayed samples is used to adjust the learning rate of the model on new samples.

[0107] Step 360: Update the parameters of the small sample continuous learning model to obtain an updated small sample continuous learning model.

[0108] It is understandable that by using the stochastic gradient descent method to update parameters during the training process of the small sample continuous learning model, it can better adapt to data of different types and characteristics and improve the flexibility, generalization ability and accuracy of the model.

[0109] In some embodiments, the calculation formula for updating the parameters of the small sample continuous learning model is as follows:

[0110] θ=θ0+γ*(θ i -θ0)P n

[0111] Among them, θ represents the updated parameters of the small sample continuous learning model, θ0 represents the initial parameters of the small sample continuous learning model, and θ i represents the parameter of the i-th gradient of the small sample continuous learning model, γ represents the outer loop learning rate, and P m Represents the updated projection operator.

[0112] It should be noted that by updating the projection operator, an orthogonal relationship is formed between the projection operator and the input space of the current task domain, and then a projection operator for a specific task domain can be maintained to constrain the optimization direction of the model parameters, thereby achieving continuous learning of small samples in multi-task domains.

[0113] Step 370: Store the new sample image of the track limit area in the memory storage space.

[0114] In some embodiments, storing the new sample image of the track limit area in the memory storage space includes:

[0115] Extract features from new sample images of the track limit area to obtain feature vectors;

[0116] The feature vector is compressed to obtain a compressed feature vector, and the compressed feature vector is stored in a memory storage space.

[0117] Optionally, new sample images of track boundary regions representing different task domains may be isolated and stored through a projection operator.

[0118] It is understandable that based on the long short-term memory mechanism, the old samples are replayed and repeatedly learned through double loops, while new data is continuously learned, which improves the generalization of the model. While continuous learning in multi-task domains, meta-learning methods are introduced, so that the model has the ability to learn with small samples and improves data utilization efficiency.

[0119] Figure 4 The second flow chart of training a small sample continuous learning model provided by an embodiment of the present invention is as follows: Figure 4As shown, in some embodiments, training a small sample continuous learning model includes:

[0120] Extract unknown category candidate regions from the new sample image of the track limit area, and determine the unknown category target region from the unknown category candidate regions;

[0121] Extract features from the target area to obtain a new feature vector;

[0122] Randomly select a memory playback sample image from the old sample image in the track limit area, perform feature extraction on the memory playback sample image, and obtain a reconstructed feature vector;

[0123] The new feature vector and the reconstructed feature vector are fused to obtain a fused feature vector;

[0124] Pooling the fused feature vector to obtain a pooled feature vector;

[0125] Based on the pooled feature vector, the obstacle recognition result corresponding to the new sample image of the track limit area is predicted to obtain the prediction result;

[0126] Based on the prediction results and the obstacle recognition result labels corresponding to the new sample images of the track limit area, the loss function value is calculated, and based on the loss function value, the parameters of the small sample continuous learning model are updated.

[0127] Among them, the small sample continuous learning model includes an encoder, a decoder, a pooling layer, a classifier, a regressor and multiple fully connected layers (FC), each FC has a weight parameter.

[0128] It should be noted that the process of training the small sample continuous learning model includes two parts: pre-training and meta-learning.

[0129] Optionally, feature extraction is performed on the new sample image of the track limit area to obtain a feature vector, and the feature vector is stored in a memory storage space.

[0130] Optionally, a frozen convolutional neural network is used to extract features from the target area to obtain a new feature vector.

[0131] Optionally, a prediction box algorithm based on edge information is used to process the new feature vector and the reconstructed feature vector to obtain a feature map with a prediction box.

[0132] Optionally, a region of interest (ROI) pooling layer is used to perform pooling processing on the feature map with the prediction box to obtain a pooled feature vector, the pooled feature vector is input into the first FC for processing, and the ROI feature vector output by the first FC is obtained, the ROI feature vector is input into the second FC for processing and then into the classifier for classification, the ROI feature vector is input into the third FC for processing and then into the regressor for classification, and finally the obstacle prediction result corresponding to the new sample image of the track limit area is obtained.

[0133] It can be understood that by continuously training the small sample continuous learning model through pre-training and meta-learning methods, the model's adaptability to new obstacles and recognition accuracy are improved through continuous learning in actual operation, which improves the performance of the track obstacle detection system in a dynamic environment and its ability to cope with new obstacles.

[0134] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rail transit obstacle recognition method based on a small sample continuous learning model, characterized in that: include: Acquire a target image, and preprocess the target image to obtain a preprocessed target image; Extracting a track limit area image from the preprocessed target image; Inputting the track limit area image into a pre-constructed small sample continuous learning model and / or an updated small sample continuous learning model, and obtaining an obstacle recognition result corresponding to the track limit area image output by the small sample continuous learning model and / or the updated small sample continuous learning model; The small sample continuous learning model is obtained by training based on old sample images of the track limit area and labels of obstacle recognition results corresponding to the old sample images of the track limit area; The updated small sample continuous learning model is obtained by continuously training the small sample continuous learning model based on the new sample image of the track limit area and the old sample image of the track limit area, as well as the obstacle recognition result label corresponding to the new sample image of the track limit area and the obstacle recognition result label corresponding to the old sample image of the track limit area, in combination with the long short-term memory mechanism; Before inputting the track limit area image into the small sample continuous learning model, the method further includes: Acquire target radar data, and preprocess the target radar data to obtain preprocessed target radar data; Based on the preprocessed target radar data, the track limit area image is classified using an open world algorithm to obtain a classification result, where the classification result includes known categories and / or unknown categories.

2. The rail transit obstacle identification method based on a small sample continuous learning model according to claim 1 is characterized in that: Before extracting the track limit area image from the preprocessed target image, the method further includes: Acquire a reference image sequence, and preprocess the reference image sequence to obtain a preprocessed reference image sequence; Extracting a track limit area reference image from the preprocessed reference image sequence; Correspondingly, extracting the track limit area image from the preprocessed target image includes: The preprocessed target image is spatially aligned with the track limit area reference image, and the track limit area image is extracted from the preprocessed target image.

3. The rail transit obstacle identification method based on a small sample continuous learning model according to claim 1 is characterized in that: The method includes: when it is determined that the classification result includes a known category, inputting the track limit area image into the pre-constructed small sample continuous learning model to obtain an obstacle recognition result corresponding to the track limit area image output by the small sample continuous learning model.

4. The rail transit obstacle identification method based on a small sample continuous learning model according to claim 1 is characterized in that: The method includes: when it is determined that the classification result includes an unknown category, inputting the track limit area image into the updated small sample continuous learning model to obtain an obstacle recognition result corresponding to the track limit area image output by the updated small sample continuous learning model.

5. The rail transit obstacle identification method based on a small sample continuous learning model according to claim 1 is characterized in that: The continuously training the small sample continuous learning model includes: Initializing parameters of the small sample continuous learning model; The old sample images of the track limit area are stored in a memory storage space, and the old sample images of the track limit area are randomly extracted from the memory storage space to obtain a memory playback sample image set; The memory playback sample image set is mixed with the track limit area new sample image to obtain a training sample image set, and a training sample image is randomly extracted from the training sample image set; In the case where the extracted training sample image is the new sample image of the track limit area, based on the new sample image of the track limit area, performing stochastic gradient descent on the parameters of the small sample continuous learning model, and updating the current projection operator; In the case where the extracted training sample image is the memory playback sample image, performing stochastic gradient descent on the parameters of the small sample continuous learning model based on the memory playback sample image; Updating the parameters of the small sample continuous learning model to obtain an updated small sample continuous learning model; The new sample image of the track limit area is stored in the memory storage space.

6. The rail transit obstacle identification method based on a small sample continuous learning model according to claim 1 is characterized in that: The continuously training the small sample continuous learning model includes: Extracting unknown category candidate regions from the new sample image of the track limit area, and determining unknown category target regions from the unknown category candidate regions; Extracting features from the target area to obtain a new feature vector; Randomly selecting a memory playback sample image from the old sample image of the track limit area, performing feature extraction on the memory playback sample image, and obtaining a reconstructed feature vector; Fusing the new feature vector and the reconstructed feature vector to obtain a fused feature vector; Pooling the fused feature vector to obtain a pooled feature vector; Based on the pooled feature vector, predicting the obstacle recognition result corresponding to the new sample image of the track limit area to obtain a prediction result; Based on the prediction result and the obstacle recognition result label corresponding to the new sample image of the track limit area, a loss function value is calculated, and based on the loss function value, the parameters of the small sample continuous learning model are updated.

7. The rail transit obstacle identification method based on a small sample continuous learning model according to claim 5 is characterized in that: The storing the new sample image of the track limit area in the memory storage space comprises: Extracting features of the new sample image of the track limit area to obtain a feature vector; The feature vector is compressed to obtain a compressed feature vector, and the compressed feature vector is stored in the memory storage space.

8. The rail transit obstacle identification method based on a small sample continuous learning model according to claim 5 is characterized in that: The expression for performing stochastic gradient descent on the parameters of the small sample continuous learning model based on the new sample image of the track limit area is as follows: i i =SGD(x i1 ,y i1 ,i i-1 ,sb); Among them, θ i represents the parameter of the i-th gradient of the small sample continuous learning model, θ i-1 represents the parameter of the i-1th gradient of the small sample continuous learning model, s represents the attention coefficient, β represents the inner loop learning rate, and x i1 represents the new sample image of the track limit area, y i1 Indicates the obstacle recognition result label corresponding to the new sample image of the track limit area; The expression for performing stochastic gradient descent on the parameters of the small sample continuous learning model based on the memory playback sample image is as follows: i i =SGD(x i2 ,y i2 ,i i-1 ,b); Among them, x i2 represents the memory playback sample image, y i2 Indicates the obstacle recognition result label corresponding to the memory playback sample image.

9. The rail transit obstacle identification method based on a small sample continuous learning model according to claim 5 is characterized in that: The calculation formula for updating the parameters of the small sample continuous learning model is as follows: θ=θ0+γ*(θ i -θ0)P m ; Wherein, θ represents the updated parameters of the small sample continuous learning model, θ0 represents the initial parameters of the small sample continuous learning model, and θ i represents the parameter of the i-th gradient of the small sample continuous learning model, γ represents the outer loop learning rate, P m Represents the updated projection operator.

Images