State space based online vectorization high precision map construction method
By using a feature storage pool based on a state-space model and a multi-directional scanning strategy, the challenges of dynamic occlusion and long-distance perception in online high-precision map construction are solved, achieving efficient and accurate high-precision map construction suitable for autonomous driving and robot perception.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2025-05-20
- Publication Date
- 2026-07-31
AI Technical Summary
Existing online high-precision map construction methods face challenges in handling dynamic occlusion and long-distance perception accuracy. A single recursive feature propagation mechanism is difficult to effectively capture long-distance temporal dependencies, and the storage and computational overhead of highly redundant temporal information is large, affecting real-time performance.
A feature storage pool based on a state-space model is adopted. Historical bird's-eye view features and instance features are selectively fused through a gating mechanism. Combined with a multi-directional scanning strategy, feature representation and instance matching are optimized to generate a high-precision map.
It improves the accuracy and adaptability of map building in dynamic environments, reduces information loss, maintains computational efficiency, is suitable for online real-time reasoning, and enhances the prediction accuracy of long-distance perception tasks.
Smart Images
Figure CN120521627B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to high-precision map online construction technology in the field of autonomous driving, specifically to an online vectorized high-precision map construction method that utilizes current camera input information and historical features, and performs temporal fusion based on a state-space model. Background Technology
[0002] High-precision maps are crucial for autonomous driving systems, providing centimeter-level road information and essential support for downstream tasks such as trajectory prediction and path planning. Traditional high-precision map construction methods typically employ offline processing based on simultaneous localization and mapping (SMR), which is computationally complex and cannot achieve real-time updates. Therefore, efficient and accurate online construction of vectorized high-precision maps has become an important research direction for autonomous driving.
[0003] In recent years, online vectorized high-precision map construction methods have gradually emerged. Mainstream methods typically use panoramic image input, extract features, convert them into a bird's-eye view representation, and then decode these features using a Transformer-based decoder to regress and obtain a vectorized map containing map elements such as road edge lines, lane dividers, and pedestrian crossings. However, in practical applications, these methods still face significant challenges in handling dynamic occlusion and maintaining long-distance perception accuracy.
[0004] To address these challenges, recent studies have found that introducing temporal modeling methods can more effectively utilize historical information, compensate for information loss caused by occlusion, and improve the prediction accuracy of distant map elements. Some studies reduce computational latency caused by the stacking of temporal information by recursively propagating hidden states. Furthermore, other studies have proposed methods to enhance the stability of temporal modeling by incorporating denoising tasks, redundant information storage, and overlapping region information.
[0005] While the aforementioned methods have improved the performance of online high-precision map construction to some extent, they still have the following shortcomings: Firstly, a single recursive feature propagation mechanism is insufficient to effectively capture long-distance temporal dependencies, resulting in significant information loss. Secondly, when using highly redundant temporal information, storage and computational overhead increases significantly, thus affecting the real-time performance of online applications. Therefore, further optimizing the temporal information fusion mechanism to achieve an efficient and accurate online vectorized high-precision map construction method remains a pressing technical challenge. Summary of the Invention
[0006] To address the problems existing in the background art, the present invention provides a novel state-space-based online vectorized high-precision map construction method for online end-to-end high-precision map construction.
[0007] The technical solution adopted in this invention is as follows:
[0008] I. A method for constructing high-precision vectorized maps based on state space
[0009] S1: Capture panoramic images using an onboard camera, extract multi-view features of the current frame from the captured panoramic images, and convert the multi-view features of the current frame into a bird's-eye view using a bird's-eye view feature encoder to generate the bird's-eye view features of the current frame.
[0010] S2: Combine the historical bird's-eye view features in the feature storage pool, optimize the current initial bird's-eye view features, obtain the optimized bird's-eye view features, and update the bird's-eye view features in the feature storage pool.
[0011] S3: Input the optimized bird's-eye view features into the map decoder to generate the initial instance features for the current frame;
[0012] S4: Combine the historical instance features in the feature storage pool, optimize the initial instance features of the current frame, obtain the optimized instance features, and update the instance features in the feature storage pool.
[0013] S5: Input the optimized instance features into the classifier and regressor respectively to predict the semantic labels and geometric shapes of map elements, thereby generating a high-precision map;
[0014] S6: Repeat S1-S5, continuously capturing panoramic images, updating the feature storage pool, and generating high-precision maps.
[0015] The feature storage pool is used to store bird's-eye view features and instance features of several historical frames and is set with a maximum number of frames.
[0016] Specifically, S2 is:
[0017] Historical bird's-eye view features are read from the feature storage pool. The read historical bird's-eye view features are fused with the initial bird's-eye view features to generate a multi-directional bird's-eye view feature sequence. Then, a state-space model based on a gating mechanism is used to selectively update the state of the multi-directional bird's-eye view feature sequence to obtain enhanced bird's-eye view feature sequences in different directions. Finally, the enhanced bird's-eye view feature sequences in different directions are integrated to obtain optimized bird's-eye view features. The optimized bird's-eye view features are added to the feature storage pool, and the bird's-eye view features in the feature storage pool are updated.
[0018] In step S2, the historical bird's-eye view features are fused with the initial bird's-eye view features to generate a multi-directional bird's-eye view feature sequence, specifically including:
[0019] First, affine transformations are performed on the historical bird's-eye view features of each frame in the feature storage pool based on the vehicle's current pose matrix to align them to the coordinate system of the current frame, thus obtaining the transformed historical bird's-eye view features of each frame. Then, all the transformed historical bird's-eye view features are concatenated with the bird's-eye view features of the current frame and then convolutional is performed to obtain an integrated feature map. Finally, the integrated feature map is converted into a multi-directional bird's-eye view feature sequence by scanning along the row or column direction, starting from different corners of the integrated feature map.
[0020] In step S2, after selectively updating the state of the multi-directional bird's-eye view feature sequence using a state-space model based on a gating mechanism, enhanced bird's-eye view feature sequences in different directions are obtained, specifically including:
[0021] The multi-directional bird's-eye view feature sequence includes original bird's-eye view feature sequences from different directions. For each original bird's-eye view feature sequence, it is first mapped to a gated signal. Then, after mapping the original bird's-eye view feature sequence to a lower dimension, a state transition is performed in the state space to obtain the transformed bird's-eye view feature sequence. Subsequently, the transformed bird's-eye view feature sequence is upgraded in dimensionality and then element-wise multiplied with the gated signal to obtain the multiplied bird's-eye view feature sequence. The multiplied bird's-eye view feature sequence in this direction is then fused with the original bird's-eye view feature sequence to obtain the enhanced bird's-eye view feature sequence in that direction. After traversing and processing the original bird's-eye view feature sequences corresponding to all directions, the enhanced bird's-eye view feature sequences in different directions are obtained.
[0022] In step S2, the enhanced bird's-eye view feature sequences from different directions are integrated to obtain the optimized bird's-eye view features, specifically including:
[0023] The enhanced bird's-eye view feature sequences from different directions are rearranged into the spatial layout of the initial bird's-eye view features according to the scanning method, and then the mean is aggregated to generate optimized bird's-eye view features.
[0024] Specifically, S4 is:
[0025] Historical instance features are read from the feature storage pool. Then, the read historical instance features are matched with the initial instance features of the current frame through inter-frame matching and multi-directional instance feature scanning to obtain a multi-directional instance feature sequence. Then, a state-space model based on a gating mechanism is used to selectively update the state of the multi-directional instance feature sequence to obtain an enhanced instance feature sequence in different directions. Finally, the enhanced instance feature sequences in different directions are integrated to obtain the optimized instance features. The optimized instance features are added to the feature storage pool and the instance features in the feature storage pool are updated.
[0026] In step S4, after inter-frame matching and fusion of the read historical instance features with the initial instance features of the current frame, a multi-directional instance feature sequence is obtained, specifically including:
[0027] The historical instance features read are matched frame by frame with the initial instance features of the current frame to obtain a multi-frame instance feature sequence with consistent order. Then, the multi-frame instance feature sequence is scanned according to the instance priority and time priority methods respectively to obtain the original instance feature sequence in the instance priority direction and the original instance feature sequence in the time priority direction, and then form a multi-directional instance feature sequence.
[0028] In step S4, after selectively updating the state of the multi-directional instance feature sequences using a state-space model based on a gating mechanism, enhanced instance feature sequences in different directions are obtained, specifically including:
[0029] The multi-directional instance feature sequence includes the original instance feature sequence in the instance-priority direction and the original instance feature sequence in the time-priority direction. For the original instance feature sequence in each direction, it is first mapped to a gating signal. Then, after mapping the original instance feature sequence in the direction to a lower dimension, a state transition is performed in the state space to obtain the transformed instance feature sequence. Subsequently, the transformed instance feature sequence is increased in dimension and then element-wise multiplied with the gating signal to obtain the dot product instance feature sequence. The dot product instance feature sequence in the direction is then fused with the original instance feature sequence to obtain the enhanced instance feature sequence in that direction. After traversing and processing the original instance feature sequences corresponding to all directions, the enhanced instance feature sequences in the instance-priority direction and the time-priority direction are obtained.
[0030] In step S4, the enhanced instance feature sequences from different directions are integrated to obtain optimized instance features, specifically including:
[0031] The enhanced feature sequences in the instance-priority direction and the time-priority direction are rearranged according to the scanning method to the spatial layout of the initial bird's-eye view feature, and then the mean is aggregated to generate the optimized bird's-eye view feature.
[0032] II. A computer device
[0033] The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method.
[0034] III. A computer-readable storage medium
[0035] A computer program is stored on a computer-readable storage medium, which, when executed by a processor, implements the steps of the method.
[0036] IV. A computer program product
[0037] The computer product includes a computer program / instructions that, when executed by a processor, implement the steps of the method.
[0038] The beneficial effects of this invention are:
[0039] This invention proposes a high-precision online map construction strategy based on a state-space model, employing a sliding window-style feature storage pool to dynamically update bird's-eye view features and instance features. Compared to traditional methods that rely solely on single-frame or short-term historical information, this invention efficiently stores and manages key features across multiple time steps, thereby effectively utilizing historical information over a longer time series. This mechanism not only improves the model's adaptability to dynamic environments but also effectively reduces information loss and enhances the accuracy of map construction under occluded conditions.
[0040] This invention introduces a gated state-space model, which selectively fuses long-term dependent information at both temporal and spatial levels. The gating mechanism precisely filters valuable historical features, avoids redundant information interference, and improves the effectiveness of information transmission. Simultaneously, the state model has linear computational complexity, making it more suitable for online real-time inference compared to methods such as Transformer. While maintaining computational efficiency, it enhances temporal modeling capabilities, enabling the system to operate efficiently in environments with limited computing power.
[0041] When constructing the feature sequence entering the state space, this invention employs a multi-directional feature scanning strategy and designs innovative scanning mechanisms for both bird's-eye view features and instance features. At the bird's-eye view feature level, the scanning strategy in different directions effectively captures multi-directional semantic information in the spatial dimension, achieving more stable temporal fusion in conjunction with the state space model. At the instance feature level, the instance-first and time-first scanning strategies optimize the consistency of cross-frame instance matching, thereby further improving the accuracy of road element prediction. This strategy can fully exploit the potential of temporal information, ensuring that the system maintains high prediction accuracy in long-distance perception tasks. Attached Figure Description
[0042] Figure 1 This is a flowchart illustrating the overall process of an online vectorized high-precision map construction method based on state space, according to an embodiment of the present invention.
[0043] Figure 2 This is a framework for an online vectorized high-precision map construction method based on state space, as described in this invention.
[0044] Figure 3 This is a schematic diagram of the sequence construction method of bird's-eye view features and instance features in an embodiment of the present invention. Detailed Implementation
[0045] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0046] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0047] like Figure 1 and Figure 2 As shown, the online vectorized high-precision map construction method based on state space proposed in this invention includes the following steps:
[0048] S1: Capture panoramic images using an onboard camera, extract multi-view features of the current frame from the captured panoramic images, and transform the multi-view features of the current frame into a bird's-eye view using a Transformer-based bird's-eye view feature encoder to generate the bird's-eye view features of the current frame.
[0049] The feature storage pool stores bird's-eye view features and instance features from several historical frames and has a maximum frame count to preserve temporal information. If the number of frames stored in the feature storage pool is less than the maximum frame count, the latest bird's-eye view features and instance features are added to the feature storage pool. If the number of frames stored in the feature storage pool is equal to the maximum frame count, the oldest frame in the feature storage pool is deleted before the latest bird's-eye view features and instance features are added to the feature storage pool.
[0050] The feature storage pool consists of two parts: the bird's-eye view feature storage pool and the instance feature storage pool. The bird's-eye view feature storage pool... The bird's-eye view features used to record historical frames are calculated using the following formula:
[0051]
[0052] Among them, F′ t-k This represents the optimized bird's-eye view feature of the tk timestamp. Instance feature storage pool. The instance features used to store historical frames are defined by the following formula:
[0053]
[0054] Among them, Q′ t-k This represents the optimized instance features at time tk. The storage pool is initialized during the first run and continuously extracts the required features during subsequent runs, updating them as the vehicle moves to ensure continuous optimization of the time-series features.
[0055] S2: Combine the historical bird's-eye view features in the feature storage pool, optimize the current initial bird's-eye view features, obtain the optimized bird's-eye view features, and update the bird's-eye view features in the feature storage pool.
[0056] S2 specifically refers to:
[0057] Historical bird's-eye view features are read from the feature storage pool. These historical features are then fused with the initial bird's-eye view features to generate a multi-directional bird's-eye view feature sequence. If historical bird's-eye view features are not available, a multi-directional bird's-eye view feature sequence is generated only based on the initial bird's-eye view features. Next, a state-space model based on a gating mechanism is used to selectively update the states of the multi-directional bird's-eye view feature sequence, resulting in enhanced bird's-eye view feature sequences for different directions. Finally, the enhanced bird's-eye view feature sequences from different directions are integrated to obtain optimized bird's-eye view features. These optimized features are then added to the feature storage pool, and the existing bird's-eye view features in the feature storage pool are updated.
[0058] In S2, the historical bird's-eye view features are fused with the initial bird's-eye view features to generate a multi-directional bird's-eye view feature sequence, specifically including:
[0059] First, affine transformations are performed on the historical bird's-eye view features of each frame in the feature storage pool based on the current pose matrix of the vehicle (i.e., the vehicle where the onboard camera is located) to align them to the coordinate system of the current frame, obtaining the transformed historical bird's-eye view features for each frame. Then, all the transformed historical bird's-eye view features are concatenated with the bird's-eye view features of the current frame, followed by a convolution operation to obtain an integrated feature map, thus integrating information from multiple frames and enhancing feature representation capabilities. Finally, as... Figure 3 As shown in (a), scanning is performed along the row or column directions starting from the four corners of the integrated feature map (top, bottom, left, right) to extract spatial correlation information and obtain a multi-directional bird's-eye view feature sequence. The multi-directional bird's-eye view feature sequence S is represented as:
[0060] S={S i |i∈{left,right,up,down}},
[0061] Where i represents the scanning direction, S i Let i represent the bird's-eye view feature sequence obtained by scanning strategy in direction i, where left, right, up, and down represent scanning strategies to the left, right, up, and down, respectively.
[0062] In S2, after selectively updating the state of the multi-directional bird's-eye view feature sequence using a state-space model based on a gating mechanism, enhanced bird's-eye view feature sequences in different directions are obtained, specifically including:
[0063] The multi-directional bird's-eye view feature sequence includes original bird's-eye view feature sequences from different directions; for each original bird's-eye view feature sequence, the original bird's-eye view feature sequence of that direction is first mapped to a gated signal V.i This is used to regulate the transmission of information flow; the formula is as follows:
[0064] V i =GELU(W v (S i ))
[0065] Where GELU() represents the Gaussian error linear unit, W v () represents the learnable parameters of a linear layer.
[0066] Then, the original bird's-eye view feature sequence in this direction is mapped to a low dimension, and a state transition is performed in the state space to capture temporal dependencies, resulting in the transformed bird's-eye view feature sequence; the formula is as follows:
[0067] U i =GELU(W u (S i )),
[0068] Y i =DSS(U i )
[0069] Among them, W u () represents the learnable parameters of a linear layer, and DSS() represents a diagonal state-space layer; U i Y represents the low-dimensional feature sequence in the i-direction after activation by the Gaussian error linear unit. i This represents the bird's-eye view feature sequence in the i-direction after transformation in the state space.
[0070] Subsequently, the transformed bird's-eye view feature sequence is up-dimensionalized and then element-wise multiplied with the gated signal to achieve adaptive information filtering, obtaining the dot-product bird's-eye view feature sequence. Then, the dot-product bird's-eye view feature sequence in this direction is fused with the original bird's-eye view feature sequence to retain key information and improve feature representation ability, obtaining the enhanced bird's-eye view feature sequence in that direction; the formula is as follows:
[0071] U i ′=W y (Y i ),
[0072] O i =W o (U i ′⊙V i )+S i
[0073] Among them, U i ′ represents the bird's-eye view feature sequence in the i-th direction after dimensionality upgrade, W y With W o Represents two learnable parameters of a linear layer, ⊙ represents the dot product operation, 0 iThis represents the enhanced bird's-eye view feature sequence in the i-direction.
[0074] After traversing and processing the original bird's-eye view feature sequences corresponding to all directions, enhanced bird's-eye view feature sequences for different directions are obtained.
[0075] In S2, the enhanced bird's-eye view feature sequences from different directions are integrated to obtain the optimized bird's-eye view features, specifically including:
[0076] The enhanced bird's-eye view feature sequences from different directions are rearranged into the spatial layout of the initial bird's-eye view features according to the scanning method, and then the mean is aggregated to generate optimized bird's-eye view features.
[0077] S3: Input the optimized bird's-eye view features into the map decoder. The map decoder interacts with the bird's-eye view features through trainable query vectors to generate the initial instance features of the current frame. The instance features are used to represent map elements.
[0078] S4: Combine the historical instance features in the feature storage pool, optimize the initial instance features of the current frame, obtain the optimized instance features, and update the instance features in the feature storage pool.
[0079] S4 specifically refers to:
[0080] Historical instance features are read from the feature storage pool. These historical features are then matched with the initial instance features of the current frame through inter-frame matching and multi-directional instance feature scanning to obtain a multi-directional instance feature sequence. If historical instance features are not present, the instance features of the current frame are used as the multi-directional feature sequence. A state-space model based on a gating mechanism is then used to selectively update the states of the multi-directional instance feature sequence to fuse historical information and enhance feature representation capabilities, resulting in enhanced instance feature sequences in different directions. Finally, the enhanced instance feature sequences in different directions are integrated to obtain optimized instance features. These optimized instance features are then added to the feature storage pool, updating the instance features in the feature storage pool.
[0081] In S4, after inter-frame matching and fusion of the read historical instance features with the initial instance features of the current frame, a multi-directional instance feature sequence is obtained, specifically including:
[0082] The historical instance features are matched frame by frame with the initial instance features of the current frame to establish inter-frame instance correspondences. The L2 distance based on instance features is used in the frame-by-frame matching process, as shown in the following formula:
[0083]
[0084] in, q represents a feature of an instance at time tk. mThis represents a feature of an instance at the current moment. Indicates the L2 distance between features. express With q m The instance matching loss.
[0085] Subsequently, the Hungarian algorithm is used to optimally allocate instances based on the matching loss, thereby obtaining a multi-frame instance feature sequence with consistent order.
[0086] like Figure 3 As shown in (b), the multi-frame instance feature sequences are then scanned in the order of instance priority and time priority to extract the instance dependencies within and between frames, obtain the original instance feature sequences in the instance priority direction and the original instance feature sequences in the time priority direction, and form a multi-directional instance feature sequence.
[0087] In S4, after selectively updating the state of the multi-directional instance feature sequences using a state-space model based on a gating mechanism, enhanced instance feature sequences in different directions are obtained, specifically including:
[0088] The multi-directional instance feature sequence includes the original instance feature sequence in the instance-priority direction and the original instance feature sequence in the time-priority direction. For the original instance feature sequence in each direction, it is first mapped to a gating signal to regulate the transmission of information flow. Then, after mapping the original instance feature sequence in the direction to a lower dimension, a state transition is performed in the state space to capture temporal dependencies, resulting in a transformed instance feature sequence. Subsequently, the transformed instance feature sequence is upgraded in dimensionality and then element-wise multiplied with the gating signal to achieve adaptive information filtering, resulting in a multiplied instance feature sequence. The multiplied instance feature sequence in this direction is then fused with the original instance feature sequence to retain key information and improve feature representation, resulting in an enhanced instance feature sequence in that direction. After traversing and processing the original instance feature sequences corresponding to all directions, the enhanced instance feature sequences in the instance-priority direction and the time-priority direction are obtained.
[0089] In S4, the enhanced instance feature sequences from different directions are integrated to obtain optimized instance features, specifically including:
[0090] The enhanced feature sequences in the instance-priority direction and the time-priority direction are rearranged according to the scanning method to the spatial layout of the initial bird's-eye view feature, and then the mean is aggregated to generate the optimized bird's-eye view feature.
[0091] S5: The optimized instance features are input into the classifier and regressor based on the multilayer perceptron network, respectively, to predict the semantic labels (including road edge lines, lane dividers, and pedestrian crossings) and geometric coordinates of map elements, thereby generating a high-precision map.
[0092] S6: Repeat S1-S5 to continuously capture surround view images, update the feature storage pool, and generate a high-precision map. Other vehicles can perform path planning and navigation based on the constructed high-precision map.
[0093] The above process outlines the common steps in both the training and inference phases of the method. Specifically, during the training phase, all network structures undergo parameter updates through training, including the bird's-eye view feature encoder, the gated state-space model for bird's-eye view features, the map decoder, the gated state-space model for instance features, the classifier, and the regressor. The existing dataset is used for training. During the training phase, additional calculations are performed to determine the polyline matching loss and classification loss between the generated map elements and the real map elements. The polyline matching loss uses a smoothed L1 loss to measure the geometric error of the polylines in the generated map, ensuring a high degree of consistency between the generated and real polylines. The classification loss employs a focus loss to enhance attention to difficult-to-classify samples, thereby suppressing the excessive influence of easily classified samples on the loss. By iteratively optimizing the model parameters, these losses are gradually reduced, ultimately improving the accuracy and consistency of the generated map.
[0094] This invention innovatively achieves dynamic management and efficient modeling of long-term historical information by introducing a sliding window-based feature storage mechanism and a gated state space model, exhibiting excellent real-time performance and multi-frame fusion capabilities. In terms of spatial modeling, the proposed multi-directional feature scanning strategy can fully exploit spatial semantics and optimize instance feature alignment, further improving the accuracy and stability of map construction in dynamic scenes and occlusion environments. While ensuring a lightweight and efficient model, this strategy effectively solves the problems of insufficient utilization of historical information and weak sequence modeling capabilities in existing methods, demonstrating strong technological advancement and practical value.
[0095] This method is applicable to fields such as autonomous driving and robot perception. This invention features end-to-end online construction of high-precision maps using time-series information, generating maps with high accuracy and robustness. This invention can be widely applied in downstream tasks of autonomous driving, such as trajectory prediction, self-localization, and path planning.
[0096] On one hand, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a method.
[0097] On one hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a method.
[0098] Finally, it should be noted that the above embodiments and descriptions are only used to illustrate the technical solutions of the present invention and not to limit it. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the disclosure of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the protection scope of the claims of the present invention.
Claims
1. A state space based online vectorization high precision map construction method, characterized in that, Includes the following steps: S1: Capture panoramic images using an onboard camera, extract multi-view features of the current frame from the captured panoramic images, and convert the multi-view features of the current frame into a bird's-eye view using a bird's-eye view feature encoder to generate the bird's-eye view features of the current frame. S2: Combine the historical bird's-eye view features in the feature storage pool, optimize the current initial bird's-eye view features, obtain the optimized bird's-eye view features, and update the bird's-eye view features in the feature storage pool. Specifically, S2 is: Historical bird's-eye view features are read from the feature storage pool. The read historical bird's-eye view features are fused with the initial bird's-eye view features to generate a multi-directional bird's-eye view feature sequence. Then, a state-space model based on a gating mechanism is used to selectively update the state of the multi-directional bird's-eye view feature sequence to obtain enhanced bird's-eye view feature sequences in different directions. Finally, the enhanced bird's-eye view feature sequences in different directions are integrated to obtain optimized bird's-eye view features. The optimized bird's-eye view features are added to the feature storage pool, and the bird's-eye view features in the feature storage pool are updated. In step S2, the historical bird's-eye view features are fused with the initial bird's-eye view features to generate a multi-directional bird's-eye view feature sequence, specifically including: First, affine transformations are performed on the historical bird's-eye view features of each frame in the feature storage pool based on the vehicle's current pose matrix to align them to the coordinate system of the current frame, thus obtaining the transformed historical bird's-eye view features of each frame. Then, all the transformed historical bird's-eye view features are concatenated with the bird's-eye view features of the current frame and then convolutional operations are performed to obtain an integrated feature map. Finally, the integrated feature map is converted into a multi-directional bird's-eye view feature sequence by scanning along the row or column direction, starting from different corners of the integrated feature map. S3: Input the optimized bird's-eye view features into the map decoder to generate the initial instance features for the current frame; S4: Combine the historical instance features in the feature storage pool, optimize the initial instance features of the current frame, obtain the optimized instance features, and update the instance features in the feature storage pool. Specifically, S4 is: Historical instance features are read from the feature storage pool. Then, the read historical instance features are matched with the initial instance features of the current frame through inter-frame matching and multi-directional instance feature scanning to obtain a multi-directional instance feature sequence. Next, a state-space model based on a gating mechanism is used to selectively update the state of the multi-directional instance feature sequence to obtain an enhanced instance feature sequence in different directions. Finally, the enhanced instance feature sequences in different directions are integrated to obtain the optimized instance features. The optimized instance features are added to the feature storage pool and the instance features in the feature storage pool are updated. S5: Input the optimized instance features into the classifier and regressor respectively to predict the semantic labels and geometric shapes of map elements, thereby generating a high-precision map; S6: Repeat S1-S5, continuously capturing panoramic images, updating the feature storage pool, and generating high-precision maps.
2. The method for constructing an online vectorized high-precision map based on state space according to claim 1, characterized in that, The feature storage pool is used to store bird's-eye view features and instance features of several historical frames and is set with a maximum number of frames.
3. The state space based online vectorization high precision map construction method according to claim 1, characterized in that, In step S2, after selectively updating the state of the multi-directional bird's-eye view feature sequence using a state-space model based on a gating mechanism, enhanced bird's-eye view feature sequences in different directions are obtained, specifically including: The multi-directional bird's-eye view feature sequence includes original bird's-eye view feature sequences from different directions. For each original bird's-eye view feature sequence, it is first mapped to a gated signal. Then, it is mapped to a lower dimension and a state transition is performed in the state space to obtain the transformed bird's-eye view feature sequence. Subsequently, the transformed bird's-eye view feature sequence is increased in dimension and then element-wise multiplied with the gated signal to obtain the multiplied bird's-eye view feature sequence. The multiplied bird's-eye view feature sequence is then fused with the original bird's-eye view feature sequence to obtain the enhanced bird's-eye view feature sequence for that direction. After traversing and processing the original bird's-eye view feature sequences corresponding to all directions, the enhanced bird's-eye view feature sequences for different directions are obtained.
4. The state space based online vectorization high-precision map construction method according to claim 2, characterized in that, In step S2, the enhanced bird's-eye view feature sequences from different directions are integrated to obtain the optimized bird's-eye view features, specifically including: The enhanced bird's-eye view feature sequences from different directions are rearranged into the spatial layout of the initial bird's-eye view features according to the scanning method, and then the mean is aggregated to generate optimized bird's-eye view features.
5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the online vectorized high-precision map construction method based on state space as described in any one of claims 1 to 4.
6. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the online vectorized high-precision map construction method based on any one of claims 1 to 4.
7. A computer program product comprising computer programs / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the state-space-based online vectorized high-precision map construction method as described in any one of claims 1 to 4.