Vehicle detection model training method and vehicle detection method
By introducing a pulsed channel attention mechanism into a pulsed dense neural network, the problem of information redundancy caused by the rapid increase in the number of channels is solved, thereby improving the accuracy and efficiency of vehicle detection.
Patent Information
- Application Number
- CN202510980795.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-16
- Publication Date
- 2025-10-17
AI Technical Summary
In the existing technology, the number of channels in the pulse dense neural network in vehicle detection increases dramatically, resulting in information redundancy and reduced detection accuracy.
In puls-dense neural networks, a pulsating channel attention mechanism is introduced. By enhancing channel weights through global average pooling, fully connected layers, and activation functions, the impact of data redundancy caused by the dramatic increase in the number of channels is reduced.
It improves the accuracy and efficiency of vehicle detection, optimizes the network structure design, and enhances the network's expressive power.
Smart Images
Figure CN120807894A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of target detection, and in particular to a vehicle detection model training method, a vehicle detection method, a vehicle detection model training device, an electronic device and a computer readable storage medium. BACKGROUND
[0002] An event camera is a kind of vision sensor inspired by the biological retina, and its working principle is different from that of a traditional frame camera which captures absolute light intensity through a fixed frame rate. The event camera adopts an asynchronous and sparse response mechanism, and only outputs an event stream when the brightness of a pixel changes. Each event contains spatial coordinates, a timestamp and polarity information. This unique perception paradigm also makes event data significantly different from frame data in terms of data processing.
[0003] In the prior art, a pulse dense neural network is generally used to process vehicle event stream data collected by an event camera to realize vehicle target detection. However, due to the multiplexing mechanism in the dense neural network, the number of channels increases sharply during network training, which causes channel information redundancy and ultimately reduces the accuracy of vehicle detection. Therefore, how to optimize the design of the detection network structure, deepen the network depth and improve the accuracy and efficiency of vehicle detection has become a key technical problem to be solved in this research. SUMMARY
[0004] The present application aims to provide a vehicle detection model training method, a vehicle detection method, a vehicle detection model training device, an electronic device and a computer readable storage medium, which are applied to the field of target detection. The method enhances the network expression capability by incorporating a pulse type channel attention mechanism into a pulse dense network structure, reduces the influence of data redundancy caused by the sharp increase in the number of channels, and improves the target detection accuracy.
[0005] To solve the above technical problems, the present application provides a vehicle detection model training method, comprising:
[0006] Obtaining vehicle event stream data captured by an event camera, and preprocessing the vehicle event stream data to obtain standard data;
[0007] Setting a pulse type channel attention mechanism in a pulse dense neural network of a vehicle detection model; the pulse type channel attention mechanism is set after a transition layer and is used to enhance the channel weight of task-related features;
[0008] Inputting the standard data into the pulse dense neural network to extract features and obtain multi-scale feature maps;
[0009] inputting the multi-scale feature map into a detection head of the vehicle detection model to obtain a vehicle detection result, updating model parameters based on a model loss determined based on the vehicle detection result until the vehicle detection model is trained to completion.
[0010] Optionally, the pulse type channel attention mechanism comprises, in sequence, a global average pooling layer, a first full connection layer, a LIF activation function, a second full connection layer, a Sigmoid activation function, and a feature rescaling layer.
[0011] Optionally, the pulse dense neural network comprises four dense blocks, and each two of the dense blocks are connected through the transition layer and the pulse type channel attention mechanism.
[0012] Optionally, the vehicle event stream data is preprocessed to obtain standard data, comprising:
[0013] The vehicle event stream data is divided according to a preset time interval to obtain a plurality of vehicle event stream data units, and each event point in the vehicle event stream data comprises an event pixel coordinate, an event timestamp, and an event polarity.
[0014] The vehicle event stream data units are normalized according to a preset space-time resolution to obtain standard vehicle event stream data units.
[0015] The standard vehicle event stream data units are subjected to voxel grid construction to obtain a dense event count tensor.
[0016] The dense event count tensor is subjected to sparse conversion to obtain a sparse event count tensor.
[0017] Optionally, the multi-scale feature map is input into the detection head of the vehicle detection model to obtain a vehicle detection result, and model parameters are updated based on a model loss determined based on the vehicle detection result, comprising:
[0018] The multi-scale feature map is predicted based on the detection head to obtain a position detection result, a category detection result, and a confidence;
[0019] Redundant detection boxes are removed from the position detection result based on a non-maximum suppression operation, and a detection box with a high confidence that meets a non-maximum condition is retained;
[0020] A positioning loss is determined based on the position detection result, and a classification loss is determined based on the category detection result;
[0021] The model parameters of the vehicle detection model are updated based on the positioning loss and the classification loss.
[0022] Optionally, the method further comprises:
[0023] An incremental learning dataset is acquired, and the vehicle detection model is updated based on the incremental learning dataset to obtain the vehicle detection model after incremental learning is completed.
[0024] To solve the above technical problems, the present application provides a vehicle detection method, comprising:
[0025] Acquiring vehicle event stream data to be detected, and preprocessing the vehicle event stream data to obtain standard data to be detected;
[0026] Inputting the standard data to be detected into a vehicle detection model to obtain a vehicle detection result;
[0027] The vehicle detection model is a model trained according to the vehicle detection model training method.
[0028] To solve the above technical problems, the present application provides a vehicle detection model training device, comprising:
[0029] A first module is configured to acquire vehicle event stream data captured by an event camera, and preprocess the vehicle event stream data to obtain standard data;
[0030] A second module is configured to set a pulse type channel attention mechanism in a pulse dense neural network of a vehicle detection model; the pulse type channel attention mechanism is set after a transition layer, and is configured to enhance channel weights of task related features;
[0031] A third module is configured to input the standard data into the pulse dense neural network to extract multi-scale feature maps;
[0032] A fourth module is configured to input the multi-scale feature maps into a detection head of the vehicle detection model to obtain a vehicle detection result, update model parameters based on a model loss determined by the vehicle detection result, and obtain the trained vehicle detection model until the training is completed.
[0033] To solve the above technical problems, the present application provides an electronic device, comprising:
[0034] A memory is configured to store a computer program;
[0035] A processor is configured to execute the computer program to implement the vehicle detection model training method or the vehicle detection method.
[0036] To solve the above technical problems, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores computer executable instructions, and the computer executable instructions are executed by a processor to implement the vehicle detection model training method or the vehicle detection method.
[0037] It can be seen that the present application obtains the vehicle event stream data captured by the event camera, pre-processes the vehicle event stream data to obtain standard data, sets a pulse type channel attention mechanism in a pulse dense neural network of a vehicle detection model, sets the pulse type channel attention mechanism after a transition layer to enhance the channel weight of the task related features, inputs the standard data into the pulse dense neural network to extract multi-scale feature maps, inputs the multi-scale feature maps into a detection head of the vehicle detection model to obtain a vehicle detection result, updates the model parameters based on the model loss determined by the vehicle detection result, and stops until a trained vehicle detection model is obtained. The vehicle event stream data to be detected is obtained, the vehicle event stream data is pre-processed to obtain standard to-be-detected data, and the standard to-be-detected data is input into the vehicle detection model to obtain a vehicle detection result. The present application enhances the network expression ability by integrating the pulse type channel attention mechanism into the pulse dense network structure, reduces the influence of data redundancy caused by the dramatic increase in the number of channels, and improves the target detection accuracy. BRIEF DESCRIPTION OF DRAWINGS
[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.
[0039] Figure 1 A flow chart of a vehicle detection model training method provided by an embodiment of the present application;
[0040] Figure 2 A structure example diagram of a pulse type channel attention mechanism provided by an embodiment of the present application;
[0041] Figure 3 A structure example diagram of a pulse dense neural network provided by an embodiment of the present application;
[0042] Figure 4 A structure block diagram of a vehicle detection model training device provided by an embodiment of the present application. DETAILED DESCRIPTION
[0043] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0044] Event cameras, as a kind of vision sensor inspired by the biological retina, work differently from traditional frame cameras that capture absolute light intensity by a fixed frame rate. Event cameras adopt an asynchronous and sparse response mechanism, and only output event streams when the brightness of the pixel changes. Each event contains spatial coordinates, a timestamp and polarity information. This unique perception paradigm also makes event data significantly different from frame data in data processing.
[0045] Compared with traditional neural networks, the event-driven characteristics, time coding ability and low power consumption of spiking neural networks make them more suitable for processing event data. However, although the spiking discrete type of spiking neural networks has a natural advantage in processing event data, it also leads to gradient discontinuity, which cannot be directly applied to back propagation. Therefore, many studies use proxy gradients such as Sigmoid functions or Atan functions to implement the back propagation of the network, but this approximation method often causes gradient vanishing or gradient explosion when processing long-time tasks, making it difficult to expand the network depth.
[0046] To solve this problem, some studies attempt to improve dense neural networks into pulse dense neural networks. Due to the feature reuse mechanism similar to residual connection in dense neural networks, incorporating the characteristics of dense neural networks into pulse neural networks can greatly improve the further expansion of network depth. However, this feature reuse mechanism can easily lead to a sharp increase in the number of channels during network training, which in turn causes channel information redundancy problems, ultimately reducing the accuracy of vehicle detection. Therefore, how to optimize the design of the detection network structure, deepen the network depth while improving the accuracy and efficiency of vehicle detection, has become a key technical problem that needs to be solved in this research.
[0047] The following combines Figure 1 , Figure 1 A flowchart of a vehicle detection model training method provided by an embodiment of the present application, which can include:
[0048] S101: Obtain the vehicle event stream data captured by the event camera, and preprocess the vehicle event stream data to obtain standard data.
[0049] The embodiment can first obtain a data set for model training, obtain vehicle event stream data through an event camera, and perform target labeling and target class labeling on the vehicle target in the event stream data. The embodiment does not limit the specific way of data labeling, which can be set based on actual application.
[0050] The embodiment is not limited to a specific type of event camera, and the vehicle event stream data can be captured in real time by a device such as a dynamic vision sensor (DAVIS, Dynamic and Active-Pixel Vision Sensor / DVS, Dynamic Vision Sensor) Each event point in the event stream data can be represented as a four-tuple asynchronous event sequence e i =(t i ,x i ,y i ,p i ), i is the event point sequence number, N is the total number of event points, e is the event, (x i ,y i ) is the pixel coordinate of the event, t i is the event timestamp, and p i belongs to {0,1} and represents the event polarity.
[0051] The embodiment can preprocess the vehicle event stream data to obtain standard data for model training. The embodiment is not limited to a specific preprocessing method. Generally, the vehicle event stream data can be divided into multiple vehicle event stream data units according to a preset time interval. Each event point in the vehicle event stream data includes an event pixel coordinate, an event timestamp, and an event polarity. The vehicle event stream data units are normalized according to a preset space-time resolution to obtain standard vehicle event stream data units. The standard vehicle event stream data units are voxel grid constructed and dense event count tensors are counted. The dense event count tensors are converted to sparse event count tensors.
[0052] The preprocessing process adopts a voxel cube coding method, and the event data is divided into time windows of a specific size in time Each time window is further subdivided into n micro time bins, and each micro time bin has a length of ∆t / n seconds. The events in each time window are finely distributed into different time units, and finally the event stream is converted into a sparse tensor form, ready for subsequent feature extraction.
[0053] S102: A pulse-type channel attention mechanism is set in the pulse dense neural network of the vehicle detection model. The pulse-type channel attention mechanism is set after the transition layer and is used to enhance the channel weight of the task-related features.
[0054] In the embodiment, a pulse-type channel attention mechanism can be set in the pulse dense neural network of the vehicle detection model to enhance the channel weight of the task-related features.
[0055] The pulse dense neural network is used for feature extraction and is composed of multiple dense blocks. A transition layer is used between every two dense blocks (DenseBlock) to perform down-sampling transition connection. The dense layers in the dense block are connected in the form of cross-channel concat, so that the network finally outputs high-level features without losing low feature information. However, the cross-connection mechanism also causes the channel number to increase rapidly with the expansion of the network depth, resulting in information redundancy.
[0056] Therefore, the embodiment introduces a pulse type channel attention mechanism (SSE, Spiking Squeeze and Excitation) after each transition layer. The vehicle event stream data is input into the pulse dense neural network after preprocessing, and is first extracted by the dense connection block. The extracted features are down-sampled by the transition layer, and the output channel information is sent to the pulse type channel attention mechanism. The pulse type channel attention mechanism processes the channel information based on the characteristics of the pulse neural network, and adaptively allocates weights to different channels, thereby effectively emphasizing key information features and suppressing non-key features.
[0057] The architecture of the pulse type channel attention mechanism in the embodiment can be as shown in Figure 2 , which includes a global average pooling layer, a first full connection layer, a LIF activation function, a second full connection layer, a Sigmoid activation function and a feature rescaling layer connected in sequence.
[0058] The pulse type channel attention mechanism processes the channel information based on the characteristics of the pulse neural network, and adaptively allocives weights to different channels. Specifically, the pulse type channel attention mechanism first compresses the spatial information by global average pooling (GAP, Global Average Pooling) on the input feature map, which can be mathematically expressed as follows:
[0059] ;
[0060] where x c represents the input feature map of the cth channel, and H and W represent the spatial dimensions of the feature map, i.e. height and width. For each channel c, the average value of the pixel values of all spatial positions (i, j) is calculated to obtain the global average value z c of all spatial position features of the channel. A lightweight fully connected layer is used to learn the channel weight, strengthen the key features and suppress the redundant information.
[0061] The overall calculation process of the pulse type channel attention mechanism can be represented as:
[0062] ;
[0063] where f is the input feature map, F is the enhanced feature output, z c is the global average value, W1 and W2 are the full connection layers for dimension reduction and dimension increase respectively, LIF (Leaky Integrate-and-Fire) is a nonlinear activation function, and σ is a Sigmoid function used to generate normalized channel weights. Finally, the weights are multiplied with the original features channel by channel through the feature re-scaling layer to realize feature calibration. The Leaky Integrate-and-Fire (LIF) model is a classic neuron model in neuroscience, which has nonlinear characteristics and is used to simulate the electrophysiological characteristics of biological neurons.
[0064] S103: input the standard data into the pulse dense neural network for feature extraction to obtain a multi-scale feature map.
[0065] In this embodiment, the preprocessed standard data can be input into the standard data pulse dense neural network for feature extraction to obtain a multi-scale feature map. The specific architecture of the pulse dense neural network is not limited in this embodiment. Generally, as shown in the figure, the pulse dense neural network includes a first dense block (Dense Block1), a first transition layer (Transition Block1), a first pulse type channel attention mechanism (SSE1), a second dense block (Dense Block2), a second transition layer (Transition Block2), a second pulse type channel attention mechanism (SSE2), a third dense block (Dense Block3), a third transition layer (Transition Block3), a third pulse type channel attention mechanism (SSE3), and a fourth dense block (Dense Block4) connected in sequence. Figure 3
[0066] In this embodiment, the preprocessed data is input into the improved pulse dense neural network for feature extraction. The network is composed of four dense blocks. A transition layer is used for down-sampling transition connection between every two dense blocks. A pulse type channel attention mechanism is introduced after each transition layer. The dense layers in the dense block are connected in the form of cross-channel concat, so that the network can output high-level features while retaining low-level feature information.
[0067] S104: input the multi-scale feature map into the detection head of the vehicle detection model to obtain a vehicle detection result. The model loss determined based on the vehicle detection result is used to update the model parameters until a trained vehicle detection model is obtained.
[0068] The embodiment does not limit the specific type of the detection head, which can generally be an SSD (Single Shot MultiBox Detector) detection head. The SSD is a single shot detector for multi-class detection. The SSD detection head is used for boundary box regression and class prediction to realize accurate positioning and recognition of the vehicle target. Through post-processing steps such as non-maximum suppression, the detection result is optimized, and the final vehicle detection result is output.
[0069] In the embodiment, the multi-scale feature map can be input into the detection head of the vehicle detection model to obtain a vehicle detection result, and the model parameters are updated based on the vehicle detection result. The model parameters are continuously updated by the vehicle event stream data until the vehicle detection model is trained.
[0070] Specifically, the multi-scale feature map is predicted based on the detection head to obtain a position detection result, a class detection result, and a confidence; redundant detection boxes are removed from the position detection result based on a non-maximum suppression operation, and a group of high-confidence detection boxes that meet the non-maximum condition are retained; a positioning loss is determined based on the position detection result, and a classification loss is determined based on the class detection result; and the model parameters of the vehicle detection model are updated based on the positioning loss and the classification loss.
[0071] Further, to ensure the stability of the model in performing the vehicle detection task, the model can be continuously updated by reinforcement learning in the embodiment, an incremental learning data set is obtained, the vehicle detection model is updated based on the incremental learning data set, and the vehicle detection model after incremental learning is completed is obtained.
[0072] Based on the above embodiment, the network expression capability is enhanced by integrating the pulse type channel attention mechanism into the pulse dense network structure, the influence of data redundancy caused by the dramatic increase in the number of channels is reduced, and the target detection accuracy is improved.
[0073] The following is a vehicle detection method provided by an embodiment of the present application, which can include:
[0074] Obtaining vehicle event stream data to be detected, and pre-processing the vehicle event stream data to obtain standard to-be-detected data;
[0075] Inputting the standard to-be-detected data into a vehicle detection model to obtain a vehicle detection result;
[0076] The vehicle detection model is a model trained according to a vehicle detection model training method.
[0077] In the embodiment, after the vehicle detection model is trained, the vehicle detection model can be deployed, such as being deployed on a cloud or a local server. The vehicle event stream data to be detected for vehicle detection is collected by an event camera, and the vehicle event stream data to be detected is transmitted to the deployment end.
[0078] Further, the vehicle event stream data is preprocessed according to a preprocessing mode of a model training process to obtain standard to-be-detected data, and the standard to-be-detected data is input into a pulse dense neural network of a vehicle detection model to obtain a multi-scale feature map through feature extraction.
[0079] A bounding box regression and class prediction are performed by using an SSD detection head to realize accurate positioning and identification of a vehicle target. Through a post-processing step such as non-maximum suppression, the detection result is optimized, and a final vehicle detection result is output.
[0080] Based on the above embodiment, the network expression capability is enhanced by incorporating a pulse type channel attention mechanism into a pulse dense network structure, the influence of data redundancy caused by a dramatic increase in the number of channels is reduced, and the target detection accuracy is improved.
[0081] The following will be described in combination with Figure 4 , Figure 4 A structural block diagram of a vehicle detection model training device provided by the embodiment of the present application can include:
[0082] A first module 100 is configured to acquire vehicle event stream data captured by an event camera, and preprocess the vehicle event stream data to obtain standard data.
[0083] A second module 200 is configured to set a pulse type channel attention mechanism in a pulse dense neural network of a vehicle detection model. The pulse type channel attention mechanism is set after a transition layer to enhance the channel weight of task-related features.
[0084] A third module 300 is configured to input the standard data into the pulse dense neural network to obtain a multi-scale feature map through feature extraction.
[0085] A fourth module 400 is configured to input the multi-scale feature map into a detection head of the vehicle detection model to obtain a vehicle detection result. Model parameters are updated based on the model loss determined by the vehicle detection result until a trained vehicle detection model is obtained.
[0086] Based on the above embodiment, the network expression capability is enhanced by incorporating a pulse type channel attention mechanism into a pulse dense network structure, the influence of data redundancy caused by a dramatic increase in the number of channels is reduced, and the target detection accuracy is improved.
[0087] Based on the above embodiment, the pulse type channel attention mechanism includes a global average pooling layer, a first full connection layer, a LIF activation function, a second full connection layer, a Sigmoid activation function, and a feature rescaling layer connected in sequence.
[0088] Based on the above embodiments, the pulse dense neural network includes four dense blocks, and each two dense blocks are connected through a transition layer and a pulse channel attention mechanism.
[0089] Based on the above embodiments, the first module 100 can include:
[0090] The first unit is configured to divide the vehicle event stream data into a plurality of vehicle event stream data units according to a preset time interval, wherein each event point in the vehicle event stream data includes an event pixel coordinate, an event timestamp, and an event polarity.
[0091] The second unit is configured to normalize the vehicle event stream data units according to a preset space-time resolution to obtain standard vehicle event stream data units.
[0092] The third unit is configured to construct a dense event count tensor by voxel grid for the standard vehicle event stream data units.
[0093] The fourth unit is configured to convert the dense event count tensor into a sparse event count tensor.
[0094] Based on the above embodiments, the fourth module 400 can include:
[0095] The fifth unit is configured to predict the multi-scale feature map based on the detection head to obtain a position detection result, a category detection result, and a confidence.
[0096] The sixth unit is configured to eliminate redundant detection boxes from the position detection result based on a non-maximum suppression operation, and retain high-confidence detection boxes that meet the non-maximum condition.
[0097] The seventh unit is configured to determine a positioning loss based on the position detection result, and determine a classification loss based on the category detection result.
[0098] The eighth unit is configured to update model parameters of the vehicle detection model based on the positioning loss and the classification loss.
[0099] Based on the above embodiments, the device can further include:
[0100] The fifth module is configured to obtain an incremental learning data set, update parameters of the vehicle detection model based on the incremental learning data set, and obtain a vehicle detection model after completion of incremental learning.
[0101] The following is a vehicle detection device provided by an embodiment of the present application, which can include:
[0102] The sixth module is configured to obtain vehicle event stream data to be detected, and pre-process the vehicle event stream data to obtain standard to-be-detected data.
[0103] The seventh module is configured to input the standard to-be-detected data into the vehicle detection model to obtain a vehicle detection result.
[0104] The vehicle detection model is a model trained by the vehicle detection model training apparatus.
[0105] Based on the above-mentioned embodiments, the present application enhances the network expression capability by integrating the pulse type channel attention mechanism in the pulse dense network structure, reduces the influence of data redundancy caused by the dramatic increase of the number of channels, and improves the target detection accuracy.
[0106] Based on the above-mentioned embodiments, the present application further provides an electronic device, which can include a memory and a processor, wherein the memory has a computer program stored therein, and the processor can realize the steps provided by the above-mentioned embodiments when calling the computer program in the memory. Of course, the device can also include various necessary network interfaces, power supplies and other components.
[0107] The present application further provides a computer readable storage medium having a computer program stored thereon, and the computer program can realize the method provided by the embodiments of the present application when executed by a terminal or a processor. The storage medium can include a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.
[0108] In this document, the terms“a,”“an,” or“the” are not intended to refer to only a singular entity, but include the general class of which a specific example can be used for illustration. The terms“a,”“an,” or“the” are used in the document to refer to one or more instances. The term“or” is used in the document to refer to a nonexclusive or, such that“A or B” means“A but not B,” or“B but not A,” and“A and B” is also included. The terms“associated with” and“associated” are used in the document to refer to one or more instances are linked together, although a specific linkage between such instances is not expressly disclosed in the specification.
Claims
1. A vehicle detection model training method, characterized in that: include: Acquire vehicle event stream data captured by an event camera, and preprocess the vehicle event stream data to obtain standard data; Setting up a spike-type channel attention mechanism in the spike-dense neural network of the vehicle detection model; The pulse channel attention mechanism is set after the transition layer to enhance the channel weights of task-related features; Inputting the standard data into the pulse dense neural network to perform feature extraction to obtain a multi-scale feature map; The multi-scale feature map is input into the detection head of the vehicle detection model to obtain a vehicle detection result, and the model parameters are updated based on the model loss determined by the vehicle detection result until the trained vehicle detection model is obtained.
2. The vehicle detection model training method according to claim 1, characterized in that: The pulse channel attention mechanism includes a global average pooling layer, a first fully connected layer, a LIF activation function, a second fully connected layer, a Sigmoid activation function and a feature recalibration layer connected in sequence.
3. The vehicle detection model training method according to claim 1, characterized in that: The pulse dense neural network includes four dense blocks, and every two dense blocks are connected through the transition layer and the pulse channel attention mechanism.
4. The vehicle detection model training method according to claim 1, characterized in that: Preprocessing the vehicle event stream data to obtain standard data includes: Dividing the vehicle event stream data into a plurality of vehicle event stream data units according to a preset time interval; each event point in the vehicle event stream data includes: event pixel coordinates, event timestamp, and event polarity; Normalizing the vehicle event stream data unit according to a preset spatiotemporal resolution to obtain a standard vehicle event stream data unit; Performing voxel grid construction on the standard vehicle event stream data unit to form a dense event count tensor; The dense event count tensor is subjected to a sparse conversion to obtain a sparse event count tensor.
5. The vehicle detection model training method according to claim 1, characterized in that: Inputting the multi-scale feature map into a detection head of the vehicle detection model to obtain a vehicle detection result, and updating model parameters based on a model loss determined based on the vehicle detection result, comprising: Predicting the multi-scale feature map based on the detection head to obtain a position detection result, a category detection result and a confidence level; Eliminating redundant detection frames from the position detection results based on a non-maximum suppression operation, and retaining high-confidence detection frames that meet non-maximum conditions; Determine a positioning loss based on the position detection result, and determine a classification loss based on the category detection result; Model parameters of the vehicle detection model are updated based on the positioning loss and the classification loss.
6. The vehicle detection model training method according to claim 1, characterized in that: Also includes: An incremental learning data set is obtained, and parameters of the vehicle detection model are updated based on the incremental learning data set to obtain the vehicle detection model after incremental learning is completed.
7. A vehicle detection method, characterized in that: include: Acquire vehicle event stream data to be detected, and preprocess the vehicle event stream data to obtain standard data to be detected; Inputting the standard data to be detected into a vehicle detection model to obtain a vehicle detection result; The vehicle detection model is a model trained according to the vehicle detection model training method according to any one of claims 1 to 6.
8. A vehicle detection model training device, characterized in that: include: The first module is used to obtain vehicle event stream data captured by an event camera and preprocess the vehicle event stream data to obtain standard data; The second module is used to set up a spike-type channel attention mechanism in the spike-dense neural network of the vehicle detection model; The pulse channel attention mechanism is set after the transition layer to enhance the channel weights of task-related features; The third module is used to input the standard data into the pulse dense neural network to extract features and obtain a multi-scale feature map; The fourth module is used to input the multi-scale feature map into the detection head of the vehicle detection model to obtain a vehicle detection result, and update the model parameters based on the model loss determined by the vehicle detection result until the trained vehicle detection model is obtained.
9. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the vehicle detection model training method according to any one of claims 1 to 6, or the vehicle detection method according to claim 7, when executing the computer program.
10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by the processor, the vehicle detection model training method according to any one of claims 1 to 6 or the vehicle detection method according to claim 7 is implemented.