Road obstacle recognition method and device

Abstract

The invention relates to a road obstacle recognition method and device. The method comprises the steps of obtaining road image data; inputting the road image data to a road obstacle recognition model to obtain a recognition result of an obstacle in the road; wherein the road obstacle recognition model comprises a first stage, a second stage, a third stage and an output layer; the first stage and the second stage are used for extracting primary features of the road image data, and the third stage is used for extracting advanced features of the road image data; wherein a jump network is also arranged between the second stage and the third stage, and the jump network is used for generating a feature vector based on the primary feature and judging whether the third stage is skipped based on the feature vector; and the output layer is used for outputting a recognition result of the obstacle in the road based on the feature vector or the advanced feature. Therefore, through judgment made by the hopping network, the flexibility and adaptability of identification can be improved, optimization processing can be carried out according to real-time data conditions, and the utilization efficiency of resources is optimized while the identification precision is ensured.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of artificial intelligence, and particularly relates to a road obstacle identification method and device. BACKGROUND

[0002] With the continuous development of technology, vehicles such as cars are becoming more and more intelligent, and to a large extent, have become an indispensable part of people's lives. But with the increase of vehicles on the road, traffic pressure is further increased, and traffic accidents occur frequently. Especially on the highway section, if the obstacles in the road cannot be cleaned in time, it may lead to accidents.

[0003] The related art mainly relies on various image processing and pattern recognition techniques for detection and identification when identifying obstacles in the road. For example, a convolutional neural network architecture is used to identify obstacles in the road. Although the related art usually has good performance when processing data, it usually requires high computing resources and processing time, and if the resources are limited or need to be processed quickly, it may not be able to obtain good identification results. SUMMARY

[0004] The present disclosure provides a road obstacle identification method and device.

[0005] According to a first aspect of the present disclosure, a road obstacle identification method is provided, the method comprising:

[0006] obtaining road image data;

[0007] inputting the road image data into a road obstacle identification model to obtain an identification result of an obstacle in the road; wherein the road obstacle identification model comprises a first stage, a second stage, a third stage and an output layer; the first stage and the second stage are used to extract primary features of the road image data, and the third stage is used to extract high-level features of the road image data; wherein a skip network is further arranged between the second stage and the third stage, the skip network is used to generate a feature vector based on the primary features, and determine whether to skip the third stage based on the feature vector; and the output layer is used to output the identification result of the obstacle in the road based on the feature vector or the high-level features.

[0008] According to a second aspect of the present disclosure, a road obstacle identification device is provided, the device comprising:

[0009] a data acquisition module configured to acquire road image data;

[0010] The obstacle recognition module is configured to input the road image data into a road obstacle recognition model to obtain an identification result of an obstacle in the road. The road obstacle recognition model comprises a first stage, a second stage, a third stage and an output layer. The first stage and the second stage are configured to extract primary features of the road image data, and the third stage is configured to extract high-level features of the road image data. A skip network is further arranged between the second stage and the third stage, and the skip network is configured to generate a feature vector based on the primary features and determine whether to skip the third stage based on the feature vector. The output layer is configured to output the identification result of the obstacle in the road based on the feature vector or the high-level features.

[0011] According to a third aspect of the present disclosure, an electronic device is provided. The electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the method described above when executing the program.

[0012] According to a fourth aspect of the present disclosure, a computer readable storage medium is provided, which stores a computer program. The program is executed by a processor to implement the method described above.

[0013] According to a fifth aspect of the present disclosure, a computer program product is provided, which comprises a computer program. The computer program is executed by a processor to implement the method described above.

[0014] The road obstacle recognition method and device provided by the embodiments of the present disclosure can obtain the identification result of the obstacle in the road by obtaining the road image data and inputting the road image data into a road obstacle recognition model. The road obstacle recognition model comprises a first stage, a second stage, a third stage and an output layer. The first stage and the second stage are configured to extract primary features of the road image data, and the third stage is configured to extract high-level features of the road image data. A skip network is further arranged between the second stage and the third stage, and the skip network is configured to generate a feature vector based on the primary features and determine whether to skip the third stage based on the feature vector. The output layer is configured to output the identification result of the obstacle in the road based on the feature vector or the high-level features. In this way, the skip network can determine whether to further process through the third stage, which can increase the flexibility and adaptability of the identification, and can optimize the processing according to the real-time data, thereby ensuring the identification accuracy and optimizing the utilization efficiency of resources. BRIEF DESCRIPTION OF DRAWINGS

[0015] In the following description of exemplary embodiments in conjunction with the accompanying drawings, more details, features and advantages of the present disclosure are disclosed, in which:

[0016] Figure 1A flowchart of a road obstacle recognition method provided for an exemplary embodiment of the present disclosure;

[0017] Figure 2 A flowchart of a road obstacle recognition method provided for an exemplary embodiment of the present disclosure;

[0018] Figure 3 A flowchart of a road obstacle recognition method provided for an exemplary embodiment of the present disclosure;

[0019] Figure 4 A functional module schematic block diagram of a road obstacle recognition apparatus provided for an exemplary embodiment of the present disclosure;

[0020] Figure 5 A structural block diagram of an electronic device provided for an exemplary embodiment of the present disclosure;

[0021] Figure 6 A structural block diagram of a computer system provided for an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0022] Embodiments of the present disclosure will be described in more detail with reference to the accompanying drawings. While the present disclosure is shown and described in connection with certain embodiments thereof, it is to be understood that the present disclosure is capable of further minor variations and modifications and can be practiced or carried out in various ways. Accordingly, it is intended that such variations and modifications not depart from the spirit and scope of the present disclosure. It should be understood that the drawings and detailed description thereto are only illustrative and are not intended to limit the present disclosure to the precise description or arrangements shown.

[0023] It should be understood that the various steps in the method embodiments of the present disclosure can be performed in different orders and / or in parallel. Additionally, the method embodiments can include additional steps and / or omit performing the steps shown. The scope of the present disclosure is not limited in this regard.

[0024] The term "comprising" and variations thereof as used herein are used inclusively, i.e., "comprising but not limited to." The term "based on" is "based at least in part on." The term "one embodiment" means "at least one embodiment." The term "another embodiment" means "at least one additional embodiment." The term "some embodiments" means "at least some embodiments." Related definitions are given throughout the detailed description. It should be noted that the concepts mentioned in the present disclosure are merely used to distinguish different apparatuses, modules or units, and are not intended to limit the order or interdependence of functions performed by these apparatuses, modules or units.

[0025] It should be noted that the modification of "one", "multiple" mentioned in the present disclosure is illustrative but not restrictive, and those skilled in the art should understand that unless otherwise explicitly indicated in the context, it should be understood as "one or more".

[0026] The names of the messages or information exchanged between the plurality of devices in the embodiments of the present disclosure are only for illustrative purposes, and are not used to limit the scope of the messages or information.

[0027] It can be understood that before using the technical solutions disclosed in the embodiments of the present disclosure, the type, use range, use scenario, etc. of the personal information involved in the present disclosure should be informed to the user and the authorization of the user should be obtained in accordance with relevant laws and regulations.

[0028] For example, in response to receiving the active request of the user, the prompt information is sent to the user to explicitly prompt the user that the operation requested to be performed will require obtaining and using the personal information of the user. Thus, the user can voluntarily choose whether to provide the personal information to the software or hardware such as electronic device, application program, server or storage medium, etc. that performs the operation of the technical solutions of the present disclosure according to the prompt information.

[0029] As an optional but non-limiting implementation manner, in response to receiving the active request of the user, the prompt information can be sent to the user in the form of a pop-up window, and the prompt information can be presented in the form of text in the pop-up window. In addition, the pop-up window can also carry a selection control for the user to select "agree" or "disagree" to provide the personal information to the electronic device. It can be understood that the above notification and obtaining of user authorization process is only illustrative, and does not limit the implementation manner of the present disclosure, and other manners meeting the relevant laws and regulations can also be applied to the implementation manner of the present disclosure.

[0030] In the related art, such as using a traditional CNN (Convolutional Neural Networks) architecture, a large amount of computing resources are often required to process complex image analysis tasks. In a resource-limited environment, such as a mobile device or a vehicle-mounted system, it can cause the system to run slowly and affect real-time performance.

[0031] For example, related technologies perform image processing and obstacle recognition through a ShuffleNet architecture. ShuffleNet is a lightweight CNN architecture designed for efficient operation on mobile devices, reducing computational cost through the use of pointwise convolutions and channel shuffle. However, the standard ShuffleNet architecture does not integrate advanced decision-making mechanisms to dynamically adjust processing flow, thus having limitations in terms of real-time performance and resource optimization. For example, implementations in related technologies typically do not include dynamic decision support systems, i.e., they routinely process every frame of image without making a judgment as to whether in-depth analysis is needed. This approach can result in unnecessary resource waste and response delays when dealing with complex or highly variable road environments.

[0032] Therefore, although the ShuffleNet architecture optimizes computational efficiency, the standard configuration may still not meet the needs of low power consumption or fast processing when dealing with high complexity or highly variable traffic environments.

[0033] As a variant of the ShuffleNet architecture, ShuffleNetV2 processes all input images through a fixed network flow, ignoring the complexity of input image content or differences in actual needs. This fixed processing flow leads to a problem where the system performs all computational steps even when faced with images of simple content. As a result, in the processing of simple or low complexity scenarios, resource utilization is low, increasing unnecessary computational burden, consuming excessive energy, and possibly causing processing delays, which is particularly pronounced on resource-constrained devices.

[0034] In addition, the lack of dynamic adjustment mechanisms makes ShuffleNetV2 unable to optimize its network structure and allocation of computational resources according to the actual image content. In real-time image processing and fast response scenarios, this fixed network structure can result in insufficient processing speed, especially in the case of dynamically changing environments or sudden events.

[0035] To address the above technical problems, the embodiments of the present disclosure improve the ShuffleNetV2 architecture and the extreme random tree gating network, which not only improves the efficiency of image processing, but also significantly optimizes resource usage through intelligent decision-making, particularly suitable for real-time and resource-constrained application scenarios. By providing a more flexible and economical solution, the embodiments can significantly improve the real-time performance and accuracy of the system, effectively addressing the deficiencies in related technologies.

[0036] The embodiments of the present disclosure can effectively solve the above technical problems by introducing a dynamic execution strategy of the SkipNet model and an Extra-Trees gated network. The embodiments can determine whether to execute some parts in the improved network according to the real-time analyzed image features. This enables the system to automatically skip unnecessary calculation layers when facing simple scenes, thereby greatly reducing resource consumption and improving processing efficiency. For complex scenes, the system can execute all layers at full capacity to ensure the accuracy of obstacle identification.

[0037] In addition, by dynamic gating decision, the embodiments of the present disclosure optimize the allocation of computing resources, and can adjust the active part of the network according to the specific needs of each scene. This optimization not only reduces energy consumption in simple scenes, but also ensures that there are enough computing resources to process high-difficulty images in complex or dangerous scenes.

[0038] Therefore, the embodiments of the present disclosure not only can significantly improve the efficiency of road obstacle identification, but also optimize the performance of the system in various operating environments, and are particularly suitable for resource-constrained or real-time requirement applications. The embodiments of the present disclosure can solve the problems of related technologies in resource utilization efficiency and network entity overload.

[0039] Specifically, the embodiments of the present disclosure provide a road obstacle identification method based on an improved ShuffleNetV2 architecture, which combines the gating idea of the SkipNet model and uses an Extra-Trees as a gating decision model. As shown in Figure 1 Figure 1 The embodiments of the present disclosure provide a flowchart of road obstacle identification, which can specifically include the following processes:

[0040] (1) Data input.

[0041] In the embodiments, image data in the current road can be collected by a camera on the vehicle or a monitoring camera in the road, or by setting a monitoring camera at a key position on the road and continuously capturing a video stream of the road situation to obtain the collected image data. The collected image data is preprocessed, and the preprocessed image data is used as the input of the data.

[0042] (2) Primary feature processing.

[0043] In the embodiments, Stage 1 and Stage 2 in ShuffleNetV2 are mainly used as the main process in the primary feature processing stage.

[0044] ​The pre-processed image data described above needs to go through a Stage 1 stage, which includes an initial convolutional layer using a 3x3 convolutional kernel, followed by a max pooling layer with the same pooling kernel and stride settings.

[0045] In the embodiment, the image data enters a basic feature extraction stage, i.e., Stage 2. This stage contains four ShuffleNet units, which perform channel splitting and channel shuffling operations inside each unit. The deep separable convolution in the Stage 2 stage can further reduce the computational cost while maintaining or improving the quality of the features. The first unit in the Stage 2 stage also uses a convolution operation with a stride of 2 to continue to reduce the size of the feature map.

[0046] It should be noted that the main role of Stage 1 and Stage 2 in ShuffleNetV2 in the embodiment is to build an efficient network structure to reduce computational complexity and maintain high accuracy. Among them, the main role of Stage 1 is to reduce the amount of calculation through the channel split operation. Specifically, Stage 1 first divides the input feature map into two parts: one part performs channel grouping operation, and the other part remains unchanged. In this way, ShuffleNetV2 can reduce the amount of calculation while maintaining feature reuse, improving the running efficiency of the model.

[0047] The main role of Stage 2 is to achieve more complex feature extraction through the stacking of multiple ShuffleNetV2 blocks. Each ShuffleNetV2 block includes channel grouping, deep separable convolution, feature rearrangement, and channel concatenation operations. These operations work together to rearrange the features and exchange information between channels, further improving the model's expression ability.

[0048] (3) Jump layer network decision.

[0049] At the transition point from Stage 2 to Stage 3 in ShuffleNetV2, a jump point is set to determine the jump decision. The jump point contains data feature extraction and dynamic decision jumping.

[0050] In the embodiment, first, in the jump point, the network further processes the data processed in the previous step, and extracts relevant features again. These features represent the state of the input data after passing through the current network layer. For example, these features can be processed by a global average pooling layer to reduce the complexity and computational requirements of subsequent processing.

[0051] The extracted features are then fed into a trained extreme random tree model, which outputs a decision indicating whether to skip the following layers. If the decision is to skip, the data is directly passed to the final output layer without going through the ShuffleNetV2 Stage 3; if the decision is not to skip, the data continues to flow through Stage 3 and then to the output layer.

[0052] The embodiments of the present disclosure combine the techniques of ShuffleNetV2 and SkipNet, and select to set a skip network at the transition point from Stage 2 to Stage 3, setting a key decision node. The core purpose of this design is to optimize the performance of the network and improve the utilization of resources through intelligent decision-making, especially in the case of complex and resource-sensitive image recognition tasks.

[0053] Stage 2 of ShuffleNetV2 is mainly responsible for basic feature extraction, and through a series of convolution and channel shuffle operations, preliminary features have been extracted from the input image. These operations are mainly to capture low-level and intermediate-level visual features such as edges and textures, which are crucial for subsequent more complex image understanding. By the end of Stage 2, a set of high-complexity feature maps has been formed, but has not yet undergone in-depth semantic analysis.

[0054] In traditional deep learning models, all input data unconditionally passes through every stage of the network. However, this processing method often leads to waste of computing resources, especially when the features are already sufficient for effective decision-making. To solve this problem, the embodiment introduces a gating mechanism of SkipNet at the transition point from Stage 2 to Stage 3. The skip network evaluates whether the feature maps obtained from Stage 2 are sufficient and whether it needs to enter Stage 3 for deeper feature processing. This evaluation is based on the output of the gating network, i.e., the skip decision, to determine whether to execute Stage 3.

[0055] This design of the embodiment enables the model to skip certain computationally intensive network layers when necessary, thereby saving resources and reducing response time. In addition, this strategy also improves the adaptability and flexibility of the network under different operating conditions and environments, enabling the model to run efficiently on resource-constrained devices, ensuring high practicality and scalability.

[0056] In the embodiment, the design of the skip network aims to optimize the computational efficiency and resource usage of the entire network, especially in resource-constrained environments. At this node, an intelligent gating mechanism is introduced, which operates based on the global average-pooled feature vector and the extreme random tree model.

[0057] Specifically, the process first involves Global Average Pooling (GAP). The purpose of this operation is to convert the multi-channel feature map output from Stage 2 into a more simplified form, thereby reducing the complexity and computational demand of subsequent processing.

[0058] Next, this feature vector is fed into an Extreme Random Tree-based gating model. This model utilizes the feature vector to decide whether to activate the complex feature extraction process of Stage 3.

[0059] The Extreme Random Tree plays a crucial role here, as it can quickly learn from these features and make decisions, with its decision function represented as:

[0060] [y = Extra-Trees(x)] (1)

[0061] Here, x is the input feature vector, and y is a binary output indicating whether to skip Stage 3. This decision mechanism allows the network to selectively skip certain computational stages without sacrificing recognition accuracy, thereby achieving higher efficiency and lower resource consumption.

[0062] By combining Global Average Pooling with Extreme Random Trees, the embodiment is able to strike a balance between high recognition performance and real-time responsiveness, making it particularly suitable for deployment in environments with limited computational resources, ensuring efficient and flexible obstacle recognition capabilities.

[0063] In the embodiments provided in the present disclosure, the Extreme Random Tree model can exhibit excellent efficiency and robustness when dealing with high-dimensional data, making it particularly suitable for real-time and dynamic decision-making environments, such as the vision processing systems of autonomous vehicles.

[0064] The core advantage of the Extreme Random Tree lies in its fast construction speed and strong generalization ability for data. This is achieved by introducing two main randomization steps during the construction of decision trees: one is to adopt a random strategy when selecting the division feature, and the other is to use random split points for each feature instead of searching for the optimal split point. This construction method not only speeds up the training process, but also increases the diversity of the model through this randomness, thereby reducing the risk of overfitting.

[0065] Another advantage of the Extreme Random Tree is its high efficiency, as once the tree is constructed, decisions can be made through a few comparisons and branches, making it very suitable for use in real-time systems. In addition, the Extreme Random Tree can effectively handle nonlinear relationships and complex interaction effects, which is particularly important when processing image data, as there are often complex spatial relationships between image features.

[0066] In the embodiments of the present disclosure, the feature vector after global average pooling provides a compressed global image information, which is used by the extreme random tree to quickly determine whether to enter Stage 3, and then perform more in-depth feature analysis.

[0067] In the embodiments provided in the present disclosure, the output of Stage 2 of ShuffleNetV2 is a feature map with rich dimensions and multiple channels. Global average pooling plays a key role in processing the output of Stage 2 of ShuffleNetV2. The purpose is to simplify the spatial information of each channel into a single numerical value, which helps to reduce the computational burden of the model in the subsequent stage and improve the processing speed. Global average pooling is achieved by averaging all pixel values in each channel, effectively compressing the features of each channel into a scalar, thereby reducing the overall feature dimension.

[0068] Specifically, if the size of each channel of the feature map is (M x N), the global average pooling operation will calculate the average of all pixel values in each channel. This can be represented by the following formula:

[0069]

[0070] where f i ( j k) represents the feature value of the kth channel at position (i, j). The result x k is the average value of the kth channel, forming a new feature vector, which includes the average values of all channels.

[0071] The global average pooling operation not only improves the efficiency of subsequent processing, but also helps to prevent overfitting by forcing the network to capture the most representative features rather than the details of specific positions. In addition, global average pooling can also simplify the output of the network, reducing the complex mapping from high-dimensional features to the final decision layer, thereby making the entire network structure more concise and efficient.

[0072] (4) Advanced feature processing.

[0073] In the embodiments, if the jump network is to perform Stage 3, the advanced feature processing step is entered, and Stage 3 stage includes 8 ShuffleNet units, each of which uses depth separable convolution and point convolution to further refine and enhance the features.

[0074] It should be noted that the main role of Stage 3 of ShuffleNetV2 is to further extract and fuse features to enhance the expression ability and accuracy of the model. In Stage 3, ShuffleNetV2 further extracts and fuses features through multiple basic units such as Shuffle modules. These modules enhance feature interaction and improve model expression ability through group convolution, channel shuffling, and element-wise addition operations.

[0075] (5) Result output.

[0076] If the skip network decides not to perform Stage 3, it enters the final result output stage.

[0077] In the embodiment, the output layer includes four ShuffleNet units, which continue to enhance feature processing through channel shuffling and depth separable convolution. At the end of the network, a global average pooling layer compresses the feature map of each channel into a single value, significantly reducing the number of parameters of the fully connected layer. A fully connected layer then maps the pooled features to the final class output, completing the obstacle recognition.

[0078] In the embodiments provided in the present disclosure, the extreme random tree plays a decisive role in the efficiency and performance optimization of the entire network. The extreme random tree decides whether to skip Stage 3 of ShuffleNetV2 by using the feature vector after global average pooling from the output of Stage 2 of ShuffleNetV2, optimizing resource use and reducing computation time. The following are the detailed steps of the extreme random tree model training.

[0079] (1) Generation of feature vector.

[0080] The embodiment will obtain a training data set including multiple images in a road. And generate a feature vector based on the image data of the training data set. Each image is processed by Stage 2 of ShuffleNetV2 to extract multi-channel feature maps. Global average pooling operation is performed on these feature maps to generate a feature vector.

[0081] Global average pooling is achieved by calculating the average of all elements in each channel, which can be achieved by the above formula (2). In this way, each image can be converted into a feature vector x containing the average values of each channel with a size of MxN.

[0082] (2) Construction of extreme random tree.

[0083] Training Extremely Randomized Trees is a process that involves building multiple independent decision trees to capture complex structures and patterns from data. This method is particularly suitable for scenarios that require fast decision-making, as it does not pursue the optimal split at each node but increases the diversity and generalization ability of the tree by introducing randomness. The following describes the training steps of Extremely Randomized Trees in detail.

[0084] The first step of training is to select a subset of features. At the beginning of building each tree, a subset of features is randomly selected from all available features. This step is crucial to increase the diversity between trees. The selection of which features will participate in the construction of the tree is random, but is usually based on a fixed proportion of all features. This approach ensures that each tree can be trained on a different subset of information, thereby increasing the overall robustness of the model.

[0085] The second step of training is to randomly select a split point. For each node in the tree, once the feature used for splitting is determined, the next step is to decide the split point on this feature. Unlike traditional decision trees, which find the optimal split point on each feature, Extremely Randomized Trees randomly select a split point on each feature. This process does not require the calculation of information gain or Gini impurity, etc., thereby greatly speeding up the construction of the tree. The specific split operation can be represented by the following formula:

[0086] [Split(S,k,t) = (x,y)∈S:x k ≤t] (3)

[0087] The third step of training is to split the node. Once the split point is determined, the samples in the node are divided into two parts according to whether they meet the split condition (x k ≤t), corresponding to the two child nodes of the tree. This process continues until a certain stopping condition is met. Common stopping conditions include: the number of samples in the node is less than a pre-set threshold, or all samples belong to the same category. These conditions help prevent overfitting and ensure that the tree does not grow too complex.

[0088] In this way, Extremely Randomized Trees introduce more randomness at each decision node, allowing the model to train on multiple different data slices, improving overall prediction performance and robustness. At the same time, this randomness also means that the model has a high tolerance for noise and outliers during training.

[0089] (3) Model training.

[0090] By repeating the above process, the Extremely Randomized Trees model is trained on the training set until all trees are constructed. Each tree is learned based on the generated feature vector, and finally reaches the ability to distinguish whether there is an obstacle or not.

[0091] (4) Usage and Validation.

[0092] After training is complete, the model can be used on new image data. For each new image, the same feature extraction and global average pooling process is performed as during training, then the trained extreme random tree model is used to predict whether to skip Stage 3. This prediction can be based on equation (1) above, where y represents the model’s predicted output and indicates whether Stage 3 needs to be performed.

[0093] This can process ensures that the model maintains high accuracy while optimizing performance and resource usage by skipping unnecessary computations.

[0094] In combination with the above embodiments, in the embodiments provided by the present disclosure, it can be specifically applied to, for example, a highway monitoring system, which uses a camera to monitor the road surface in real time in order to detect and respond to various obstacles such as debris flying from the outside, items left by vehicles, damage to the road surface (such as road cracks or collapse, etc.), and congestion caused by traffic accidents, etc.

[0095] Firstly, the monitoring system relies on high-quality camera equipment that needs to be able to clearly capture road images under various lighting and weather conditions. The camera should have high resolution and sufficient frame rate to capture objects moving at high speed while ensuring the clarity of the image, which is the basis for the detection accuracy of the system.

[0096] Secondly, the monitoring system needs a backend processing unit, which can be implemented through a server installed near the road or using cloud computing resources. These processing units need to have sufficient computing power to run the methods involved in the embodiments of the present disclosure to process video streams transmitted from the camera in real time.

[0097] In addition, the implementation of the system is best able to coordinate with the local traffic management system in order to take quick action when serious obstacles are detected. This includes integration with emergency response systems, or direct linkage with traffic signal control systems to adjust traffic flow and alleviate traffic pressure in specific areas.

[0098] Specifically, as shown in Figure 2 , a flowchart of road obstacle identification provided by another embodiment of the present disclosure is shown. Specifically, it can include the following steps: Figure 2

[0099] (1) Data collection and preliminary processing.

[0100] ​Image data in the current road can be collected by cameras on vehicles or monitoring cameras in the road. For example, during the driving of the vehicle, image data in the current road can be collected in real time by the cameras arranged on the vehicle. Alternatively, image data collected by setting monitoring cameras at key positions on the road and continuously capturing video streams of road conditions. These key positions usually include intersections, accident-prone areas, and other important traffic nodes to ensure that the system can comprehensively monitor and record various conditions on the road.

[0101] The image data can be pre-processed before being delivered to the processing center to improve the availability of the data. This includes automatically adjusting the sharpness of the image to cope with different ambient light conditions; applying advanced noise reduction techniques to eliminate random noise in the image, which is particularly important at night or in bad weather conditions; and color correction to ensure the authenticity and consistency of the image color, so that the subsequent feature recognition is more accurate. These preprocessing operations ensure the quality of the input data and improve the sensitivity and accuracy of the system for obstacle recognition.

[0102] After the image preprocessing is completed, that is, the image data is sharpened and denoised, and color correction is performed, the preprocessed image data is then converted into a format suitable for deep learning model processing. The video stream is segmented into single frames or small batches of frames, the size of the image is adjusted to adapt to the input requirements of the model, and the image file is converted into a specific data type and structure. After the above processing, it can be used for subsequent feature extraction and analysis.

[0103] (2) Feature extraction and jump network decision.

[0104] In the embodiment, the image data described above can be received for initial processing, which has been optimized to meet the needs of the deep learning model. On this basis, the Stage 2 in the improved ShuffleNetV2 architecture is used to focus on efficiently extracting deep features from these images. This stage uses its advanced network design to extract basic but key visual information such as edges, textures and shapes from images through grouped convolution and channel shuffle techniques.

[0105] After completing the feature extraction, the system uses global average pooling processing. The purpose of global average pooling is to reduce the dimension and complexity of the data while retaining the key information. The feature vector after global average pooling is sent into the jump network. The core function of the jump network is to evaluate these feature vectors and make decisions based on the extreme random tree model. Decide whether to enter Stage 3 in the ShuffleNetV2 architecture for deeper feature analysis or jump to the last stage. This decision is a key link for the system to optimize the calculation efficiency and response speed.

[0106] The end of this step marks the completion of the gating network's decision, outputting whether to activate Stage 3's results. This decision ensures that the system maintains high recognition accuracy while optimizing resource usage and processing speed to the maximum.

[0107] (3) Advanced Feature Processing.

[0108] Advanced Feature Processing is the stage in the system where in-depth feature analysis is conducted. The main goal is to extract more complex and abstract features from the image, laying a solid foundation for accurate obstacle recognition.

[0109] The decision to start this step depends on the outcome of the skip network's decision. When the skip network decides not to skip Stage 3, the image data is introduced to Stage 3, initiating a deeper feature extraction process. This signifies that the system needs to conduct more detailed analysis to ensure the accuracy of the recognition results, especially in complex or ambiguous scenarios.

[0110] The end of this step marks that the image data has been thoroughly analyzed, and all critical information has been captured by the system and is ready for decision output, such as the type of obstacle recognized, its location, and its potential impact on traffic.

[0111] (4) Output and Response.

[0112] Output and Response is the final stage in the road obstacle recognition system, tasked with classifying and identifying obstacles based on the features extracted and processed in previous stages, and then generating the necessary output and taking appropriate response measures.

[0113] This step has two sources of data. One is the advanced features extracted from Stage 3; the other is the feature data received from the skip network. These feature data include all necessary information about the obstacle, such as its type, size, location, and potential impact, providing a basis for the final decision.

[0114] In this stage, feature data is analyzed to determine the type of obstacle present in the image. This may include simple objects such as stones on the road, discarded items, to more complex situations such as road blockages or road damage caused by accidents. The classification process relies on well-trained algorithms that can accurately identify various obstacles from features and judge their potential impact on traffic.

[0115] Once the obstacle is correctly identified, the system generates a corresponding output and triggers a response measure. This includes automatically sending an alert to the traffic management center, providing detailed information about the obstacle to take quick action; activating a road display warning, such as an electronic sign or a variable message sign, to warn the driver of the upcoming danger; or in more urgent cases, directly interfacing with the emergency service system to ensure a quick accident response and handling.

[0116] The embodiments of the present disclosure are based on the improved ShuffleNetV2 architecture, and specific improvements are made by adopting the ShuffleNetV2 architecture, including optimizing the network level and parameters to enhance the ability to process high complexity image data under low resource consumption. The optimization of channel splitting and shuffling operations makes the network more suitable for real-time processing of a large amount of monitoring video data. And the embodiments allow the system to automatically adjust the processing according to the complexity of the current image features through the skip network strategy, which can increase the flexibility and adaptability of the system, and optimize the processing path according to the real-time data situation, thereby maximizing the effective use of resources while ensuring recognition accuracy.

[0117] Therefore, the embodiments of the present disclosure can greatly improve the computational efficiency of the network by adopting the improved ShuffleNetV2 architecture, while maintaining the lightweight of the traditional deep learning model. This can achieve high-quality image processing and feature recognition with lower computational resource consumption in a resource-constrained environment, solving the contradiction between processing speed and resource consumption in related technologies.

[0118] And the embodiments introduce an extreme random tree as a decision model of the gating network, which can selectively skip some network layers based on real-time analysis of data requirements and the current network state, thereby optimizing the overall computational efficiency. This strategy not only reduces unnecessary computational burden, but also speeds up the response time of the system, especially when fast recognition and processing of complex traffic situations are required, which can significantly improve the practicality and efficiency of the system.

[0119] Based on the above embodiments, the embodiments of the present disclosure first provide a road obstacle identification method, as shown in Figure 3 The method can include the following steps:

[0120] In step S310, road image data is obtained.

[0121] In the embodiments, the image data in the current road can be collected by a camera on the vehicle or a monitoring camera in the road. For example, during the driving of the vehicle, the image data in the current road can be collected in real time by the camera arranged on the vehicle.

[0122] In an embodiment, the image data collected can also be obtained by setting up monitoring cameras at key locations on the road and continuously capturing video streams of the road conditions. These key locations typically include intersections, accident-prone areas, and other important traffic nodes to ensure that the system can comprehensively monitor and record various situations on the road.

[0123] In an embodiment, the collected image data can also be pre-processed. For example, by obtaining road images in the target road in real time and pre-processing the road images to obtain road image data. The pre-processing includes clarity processing, denoising processing or color correction processing.

[0124] In an embodiment, the pre-processing process can include adjusting the clarity of the image to cope with different environmental lighting conditions. Advanced denoising techniques can also be applied to eliminate random noise in the image, which is particularly important at night or in adverse weather conditions. And the image can also be color corrected to ensure the authenticity and consistency of the image color, so that the subsequent feature recognition is more accurate. These preprocessing operations ensure the quality of the input data, and improve the sensitivity and accuracy of the system in identifying obstacles.

[0125] In step S320, the road image data is input into the road obstacle recognition model to obtain the recognition result of the obstacles in the road.

[0126] The road obstacle recognition model includes a first stage, a second stage, a third stage and an output layer; the first stage and the second stage are used to extract the primary features of the road image data, and the third stage is used to extract the high-level features of the road image data; wherein a skip network is also provided between the second stage and the third stage, the skip network is used to generate a feature vector based on the primary features, and determine whether to skip the third stage based on the feature vector; the output layer is used to output the recognition result of the obstacles in the road based on the feature vector or the high-level features.

[0127] In an embodiment, the road barrier recognition model can be the improved ShuffleNetV2 in the above embodiments, the first stage is Stage 1, the second stage is Stage 2, and the third stage is Stage 3. By setting a skip network at the transition point between Stage 2 and Stage 3, a key decision node is set, so that the performance of the network can be optimized and the utilization of resources can be improved through the decision of the skip network, especially in the case of complex and resource-sensitive image recognition tasks.

[0128] The jump network evaluates whether the feature map obtained from Stage 2 is sufficient to obtain the identification result. If so, the primary features obtained from Stage 2 are used to generate a feature vector, which is transmitted to the output layer to output the identification result of the obstacle in the road. Otherwise, the processing of Stage 3 is still needed to extract more advanced features, which are transmitted to the output layer to output the identification result of the obstacle in the road.

[0129] In the embodiment, when the feature vector is generated based on the primary features and whether to skip the third stage is determined based on the feature vector, the global average pooling processing can be performed on the primary features by the jump network to obtain the feature vector. The feature vector is input into the extreme random tree model in the jump network to output the output result of whether to skip the third stage.

[0130] In the embodiment, the global average pooling processing can convert the multi-channel feature map output by Stage 2 into a more simplified form, thereby reducing the complexity and computational requirements of subsequent processing.

[0131] The road obstacle identification method provided by the embodiment of the present disclosure obtains the identification result of the obstacle in the road by acquiring the road image data and inputting the road image data into the road obstacle identification model. The road obstacle identification model includes a first stage, a second stage, a third stage, and an output layer. The first stage and the second stage are used to extract primary features of the road image data, and the third stage is used to extract advanced features of the road image data. The jump network is arranged between the second stage and the third stage, which is used to generate a feature vector based on the primary features and determine whether to skip the third stage based on the feature vector. The output layer is used to output the identification result of the obstacle in the road based on the feature vector or the advanced features. The jump network makes a judgment on whether to further process by the third stage, which can increase the flexibility and adaptability of the identification and optimize the processing according to the real-time data, thereby ensuring the identification accuracy and optimizing the utilization efficiency of resources.

[0132] Based on the above embodiment, when the output layer outputs the identification result of the obstacle in the road based on the feature vector or the advanced features, the output layer is used to output the identification result of the obstacle in the road based on the feature vector when the output result is to skip the third stage. Or, when the output result is not to skip the third stage, the advanced features of the road image data are extracted based on the feature vector by the third stage, and the output layer is used to output the identification result of the obstacle in the road based on the advanced features.

[0133] In the embodiment, the feature vector after global average pooling can provide compressed global image information. Based on this global image information, the extreme random tree can quickly determine whether it needs to enter Stage 3, that is, whether it needs to perform more in-depth feature analysis through Stage 3.

[0134] In this embodiment, when constructing the extreme random tree model in the skip network, training data can be obtained and the extreme random tree model can be constructed based on this training data. The training data includes feature vectors generated from road images of multiple roads. For details, please refer to the description of the extreme random tree model training process in the above embodiments; it will not be repeated here.

[0135] In the embodiments provided in this disclosure, the category information of obstacles in the road can be obtained based on the above identification results. Road warning information can then be generated based on this category information.

[0136] For example, effective measures can be taken promptly based on the category information of obstacles on the road. For instance, when the category information is "road collapse," a road repair warning can be generated in a timely manner to facilitate road repair. When the category information is "tire" or other road obstacles, an obstacle removal warning can be generated to facilitate timely removal of the obstacles and prevent traffic accidents.

[0137] By dividing each function into corresponding functional modules, this disclosure provides a road obstacle recognition device, which can be a server, a terminal, or a chip applied to a server. Figure 4 This is a schematic block diagram illustrating the functional modules of a road obstacle recognition device provided as an exemplary embodiment of this disclosure. Figure 4 As shown, the road obstacle recognition device includes:

[0138] Data acquisition module 10 is used to acquire road image data;

[0139] The obstacle recognition module 20 is used to input the road image data into a road obstacle recognition model to obtain the recognition result of obstacles in the road; wherein, the road obstacle recognition model includes a first stage, a second stage, a third stage, and an output layer; the first stage and the second stage are used to extract primary features of the road image data, and the third stage is used to extract high-level features of the road image data; wherein, a skip network is further provided between the second stage and the third stage, the skip network is used to generate feature vectors based on the primary features, and determine whether to skip the third stage based on the feature vectors; the output layer outputs the recognition result of obstacles in the road based on the feature vectors or the high-level features.

[0140] In another embodiment provided in this disclosure, the obstacle recognition module is specifically used for:

[0141] performing global average pooling processing on the primary features through the jump network to obtain a feature vector;

[0142] inputting the feature vector into an extreme random tree model in the jump network, and outputting an output result of whether to skip the third stage.

[0143] In another embodiment provided by the present disclosure, the obstacle recognition module is further configured to:

[0144] in a case where the output result is to skip the third stage, the output layer is configured to output a recognition result of an obstacle in the road based on the feature vector;

[0145] in a case where the output result is not to skip the third stage, the third stage is configured to extract high-level features in the road image data based on the feature vector, and the output layer is configured to output the recognition result of the obstacle in the road based on the high-level features.

[0146] In another embodiment provided by the present disclosure, the device further comprises a training module configured to:

[0147] obtain training data, the training data comprising feature vectors generated based on road images in a plurality of roads;

[0148] construct the extreme random tree model based on the training data.

[0149] In another embodiment provided by the present disclosure, the data obtaining module is configured to:

[0150] obtain road images in a target road in real time;

[0151] perform preprocessing on the road images to obtain the road image data, wherein the preprocessing comprises definition processing, denoising processing, or color correction processing.

[0152] In another embodiment provided by the present disclosure, the device further comprises:

[0153] a category information obtaining module configured to obtain category information of an obstacle in the road based on the recognition result;

[0154] a warning module configured to generate road warning information based on the category information.

[0155] The road obstacle identification device provided by the embodiments of the present disclosure obtains the identification result of the obstacle in the road by obtaining the road image data and inputting the road image data into the road obstacle identification model. The road obstacle identification model includes a first stage, a second stage, a third stage and an output layer. The first stage and the second stage are used to extract the primary features of the road image data, and the third stage is used to extract the advanced features of the road image data. A skip network is further arranged between the second stage and the third stage, which is used to generate a feature vector based on the primary features and determine whether to skip the third stage based on the feature vector. The output layer is used to output the identification result of the obstacle in the road based on the feature vector or the advanced features. In this way, the skip network makes a judgment on whether to further process through the third stage, which can increase the flexibility and adaptability of the identification, and can optimize the processing according to the real-time data, thereby ensuring the identification accuracy and optimizing the utilization efficiency of resources.

[0156] The embodiments of the present disclosure also provide an electronic device, including: at least one processor; a memory for storing instructions executable by the at least one processor; wherein the at least one processor is configured to execute the instructions to implement the above-mentioned method disclosed by the embodiments of the present disclosure.

[0157] Figure 5 The structural schematic diagram of the electronic device provided by an exemplary embodiment of the present disclosure is shown in FIG. 18. As shown in the figure, the electronic device 1800 includes at least one processor 1801 and a memory 1802 coupled to the processor 1801, and the processor 1801 can execute the corresponding steps in the above-mentioned method disclosed by the embodiments of the present disclosure. Figure 5

[0158] ​The processor 1801 can also be referred to as a central processing unit (CPU), which can be an integrated circuit for processing signals. Each of the steps of the methods disclosed in the embodiments of the present disclosure can be completed by an integrated logic circuit of hardware or an instruction in the form of software in the processor 1801. The processor 1801 can be a general-purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the methods disclosed in conjunction with the embodiments of the present disclosure can be directly embodied as a hardware coding processor for execution, or a combination of hardware and software modules in the coding processor for execution. The software module can be located in a memory 1802, such as a random access memory, a flash memory, a read only memory, a programmable read only memory, an electrically erasable programmable memory, a register, and other mature storage mediums in the art. The processor 1801 reads information in the memory 1802 and completes the steps of the methods in conjunction with the hardware.

[0159] In addition, various operations / processes according to the present disclosure, when implemented by software and / or firmware, can be loaded from a storage medium or a network into a computer system having a dedicated hardware structure, such as the computer system 1900 shown. Figure 6 The computer system 1900 shown is intended to represent various forms of digital electronic computer equipment, such as a laptop computer, a desktop computer, a workstation, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computer equipment. The electronic equipment can also represent various forms of mobile devices, such as a personal digital processor, a cellular phone, a smart phone, a wearable device, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are merely examples, and are not intended to limit the implementations described and / or claimed in this document. Figure 6 A structural block diagram of a computer system provided for an exemplary embodiment of the present disclosure.

[0160] The computer system 1900 is intended to represent various forms of digital electronic computer equipment, such as a laptop computer, a desktop computer, a workstation, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computer equipment. The electronic equipment can also represent various forms of mobile devices, such as a personal digital processor, a cellular phone, a smart phone, a wearable device, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are merely examples, and are not intended to limit the implementations described and / or claimed in this document.

[0161] As shown in FIG. 19, the computer system 1900 includes one or more processors, such as a processor 1904. The processor 1904 is connected to a communication infrastructure 1902, such as a bus, cross-over bar, or network. Figure 6As shown, the computer system 1900 includes a computing unit 1901 that can perform various appropriate actions and processes in accordance with a computer program stored in a read-only memory (ROM) 1902 or a computer program loaded into a random access memory (RAM) 1903 from a storage unit 1908. Various programs and data required for the operation of the computer system 1900 can also be stored in the RAM 1903. The computing unit 1901, the ROM 1902, and the RAM 1903 are connected to each other through a bus 1904. An input / output (I / O) interface 1905 is also connected to the bus 1904.

[0162] Various components in the computer system 1900 are connected to the I / O interface 1905, including an input unit 1906, an output unit 1907, a storage unit 1908, and a communication unit 1909. The input unit 1906 can be any type of device that can input information to the computer system 1900, and can receive inputted digital or character information, and generate key signal inputs related to user settings and / or function controls of the electronic device. The output unit 1907 can be any type of device that can present information, and can include, but is not limited to, a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 1908 can include, but is not limited to, a magnetic disk, an optical disk. The communication unit 1909 allows the computer system 1900 to exchange information / data with other devices through a network such as the Internet, and can include, but is not limited to, a modem, a network card, an infrared communication device, a wireless communication transceiver, and / or a chipset, such as a Bluetooth™ device, a WiFi device, a WiMax device, a cellular communication device, and / or the like.

[0163] The computing unit 1901 can be various general and / or special purpose processing components with processing and computing capabilities. Some examples of the computing unit 1901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 1901 performs various methods and processes described above. For example, in some embodiments, the above-described methods disclosed by embodiments of the present disclosure can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 1908. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device via the ROM 1902 and / or the communication unit 1909. In some embodiments, the computing unit 1901 can be configured to perform the above-described methods disclosed by embodiments of the present disclosure by any other appropriate means (e.g., by means of firmware).

[0164] The embodiment of the present disclosure further provides a computer readable storage medium, wherein when instructions in the computer readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the above method disclosed by the embodiment of the present disclosure.

[0165] The computer readable storage medium in the embodiment of the present disclosure can be a tangible medium, which can contain or store programs for use by or in connection with an instruction execution system, apparatus or device. The above computer readable storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or any suitable combination of the above. More specifically, the above computer readable storage medium can include one or more wire-based electrical connections, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0166] The above computer readable medium can be included in the above electronic device; or can exist separately without being assembled into the electronic device.

[0167] The embodiment of the present disclosure further provides a computer program product, comprising a computer program, wherein the computer program is executed by a processor to implement the above method disclosed by the embodiment of the present disclosure.

[0168] In the embodiments of the present disclosure, computer program code for carrying out operations of the present disclosure can be written in one or more programming languages or combinations of the same, including object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" language or similar programming languages. Program code can be executed entirely on a user computer, partially on a user computer, as an independent software package, partially on a user computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user computer through any kind of network, including local area network (LAN) or wide area network (WAN), or can be connected to an external computer.

[0169] The flow diagrams and the block diagrams in the drawings are illustrations of architectures, functionalities, and operations of possible implementations of systems, methods, and computer program products according to various embodiments of present disclosure. In this regard, each block in the flow diagrams or block diagrams can represent a module, a segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that in some alternative implementations, the functions noted in the block can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flow diagrams, and combinations thereof, can be implemented by special purpose hardware-based systems that perform the specified functions or operations, or combinations of special purpose hardware and computer instructions.

[0170] The modules, components or units described in the embodiments of the present disclosure can be implemented by software or by hardware. In some cases, the name of the module, component or unit does not constitute a limitation on the module, component or unit itself.

[0171] The functions described in this specification can be implemented in part or in whole by one or more hardware logic components. For example, and without limitation, illustrative types of hardware logic components that can be used include Field-programmable Gate Arrays (FPGAs), Program-specific Integrated Circuits (ASICs), Program-specific Standard Products (ASSPs), System-on-a-chip systems (SOCs), Complex Programmable Logic Devices (CPLDs), etc.

[0172] The above description is merely illustrative of the embodiments of the present disclosure and the principles of the technology involved. It is understood that the disclosure scope of the present disclosure is not limited to the technical solutions formed by the specific combinations of the above technical features, and should also cover other technical solutions formed by the combinations of the above technical features or equivalent features thereof without departing from the above disclosure concept. For example, the technical solutions formed by the mutual replacement of the above features and the technical features disclosed in the present disclosure (but not limited to) having similar functions.

[0173] Although some specific embodiments of the present disclosure have been described in detail by way of examples, one skilled in the art should understand that the above examples are only for illustration, and are not intended to limit the scope of the present disclosure. One skilled in the art should understand that the above embodiments can be modified without departing from the scope and spirit of the present disclosure. The scope of the present disclosure is defined by the appended claims.

Claims

1. A road obstacle recognition method characterized by comprising: The method comprises: acquiring road image data; inputting the road image data into a road obstacle recognition model to obtain a recognition result of an obstacle in a road; wherein the road obstacle recognition model comprises a first stage, a second stage, a third stage and an output layer; the first stage and the second stage are used to extract primary features of the road image data, and the third stage is used to extract high-level features of the road image data; wherein a skip network is further arranged between the second stage and the third stage, the skip network is used to generate a feature vector based on the primary features, and determine whether to skip the third stage based on the feature vector; and the output layer is used to output the recognition result of the obstacle in the road based on the feature vector or the high-level features.

2. The method of claim 1, wherein, The method further comprises: generating the feature vector based on the primary features, and determining whether to skip the third stage based on the feature vector, comprises: performing global average pooling processing on the primary features through the skip network to obtain a feature vector; 3. The method of claim 2, wherein, inputting the feature vector into an extreme random tree model in the skip network to output an output result of whether to skip the third stage. The output layer is used to output the recognition result of the obstacle in the road based on the feature vector or the high-level features, comprising: in the case that the output result is to skip the third stage, the output layer is used to output the recognition result of the obstacle in the road based on the feature vector; 4. The method of claim 2, wherein, or, in the case that the output result is not to skip the third stage, extracting high-level features in the road image data based on the feature vector through the third stage, and the output layer is used to output the recognition result of the obstacle in the road based on the high-level features. The method further comprises: acquiring training data, the training data comprising feature vectors generated based on road images in a plurality of roads; 5. The method of claim 1, wherein, constructing the extreme random tree model based on the training data. The method further comprises: real-time acquiring road images in a target road; 6. The method of claim 1, wherein, preprocessing the road images to obtain the road image data. The method further comprises: based on the recognition result, acquiring category information of the obstacle in the road; 7. A road obstacle recognition apparatus characterized by comprising: generating road warning information based on the category information. The device comprises: a data acquisition module configured to acquire road image data; 8. An electronic device, comprising: an obstacle recognition module configured to input the road image data into a road obstacle recognition model to obtain a recognition result of an obstacle in a road; wherein the road obstacle recognition model comprises a first stage, a second stage, a third stage and an output layer; the first stage and the second stage are used to extract primary features of the road image data, and the third stage is used to extract high-level features of the road image data; wherein a skip network is further arranged between the second stage and the third stage, the skip network is used to generate a feature vector based on the primary features, and determine whether to skip the third stage based on the feature vector; and the output layer is used to output the recognition result of the obstacle in the road based on the feature vector or the high-level features. comprises: at least one processor; a memory for storing the instructions executable by the at least one processor; wherein the at least one processor is configured to execute the instructions to implement the method of any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium, when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform the method of any one of claims 1-6.

10. A computer program product, characterised in that, A computer program product, which comprises a computer program that, when executed by a processor, implements the method of any one of claims 1-6.

Images