Online LiDAR moving target segmentation method based on Haar wavelet enhancement

Abstract

The invention discloses an online LiDAR moving target segmentation method based on Haar wavelet enhancement, and belongs to the technical field of target segmentation. The problem that an existing moving target online segmentation method is low in precision is solved. According to the method, a residual multi-layer perceptron structure is adopted, the residual multi-layer perceptron structure can reduce information loss caused by depth reduction by introducing a residual mechanism, the mapping quality from the point cloud to the BEV feature map is improved, and spatial information expression is enhanced. By designing the wavelet fusion block, multi-scale feature modeling can be enhanced based on Haar wavelet transform, information expression of a BEV feature map in the horizontal direction, the vertical direction and the diagonal direction is enhanced, and the capturing capacity of a moving target is improved. The optimized coding-decoding architecture can reduce the calculation overhead while improving the segmentation precision, and meets the real-time requirements of online tasks. Through point cloud feature extraction, feature fusion modeling and segmentation network design, segmentation precision and real-time performance are improved. The method can be applied to moving target segmentation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of target segmentation, and particularly relates to an online LiDAR moving target segmentation method based on Haar wavelet enhancement. BACKGROUND

[0002] As a core sensor of environmental perception, the laser radar technology plays a key role in intelligent perception fields such as automatic driving systems and robot navigation, can well restore the details of the space environment around the vehicle, and has a low influence coefficient on light and climate. The moving target segmentation task based on the laser radar aims to accurately distinguish dynamic targets from static backgrounds. However, due to the discreteness, sparsity and disorder of point cloud data itself, the existing segmentation methods still face many challenges in feature expression, computational efficiency and generalization ability.

[0003] To address the above challenges, online moving object segmentation (MOS) task emerges to realize real-time identification and segmentation of dynamic objects in LiDAR point cloud without relying on future frames, so as to meet the requirements of continuous perception, path adjustment and emergency obstacle avoidance for the response speed and stability of key functions of autonomous driving. Current mainstream LiDAR-MOS methods mainly include point cloud / voxel processing, range view projection and bird's eye view projection. Point cloud / voxel method directly extracts features in three-dimensional space, which completely retains the spatial structure information. For example, PointNet (Qi C R, Yi L, Su H, et al. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space[J]. 2017. DOI: 10.48550 / arXiv. 1706.02413.) series enhances the representation ability of point cloud through hierarchical sampling, and 4DMOS (Mersch B, Chen X, Vizzo I, et al. Receding moving object segmentation in 3d lidar data using sparse 4d convolutions[J]. IEEE Robotics and Automation Letters, 2022, 7(3): 7503-7510.) models the motion trajectory based on time series information, thereby improving the identification effect of dynamic objects. Although such methods can effectively capture motion features, the high computational complexity limits the real-time processing capability, especially on the embedded platform of autonomous driving. RV projection method converts point cloud into two-dimensional depth view, and then extracts its features by convolutional neural network to realize the segmentation of moving objects. For example, RVMOS (Kim J, Woo J, Im S. Rvmos: Range-view moving object segmentation leveraged by semantic and motion features[J]. IEEE Robotics and Automation Letters, 2022, 7(3): 8044-8051.) extracts motion features by residual calculation on continuous multiple frames of point cloud; MF-MOS decouples spatial structure and motion features on this basis, enhancing the generalization performance. Such methods are favored due to high computational efficiency and simple network structure, but they still have obvious limitations. In the target occlusion scene, they are prone to large area misjudgment, and the identification ability of low-speed targets is weak, so the identification accuracy in complex scenes is still low.The BEV projection method extracts spatial features through gridded mapping and combines them with a two-dimensional segmentation network to complete the MOS task. For example, LiMoSeg (Mohapatra S, Yogamani S, Kumar VR, et al. LiDAR-BEVMTN: Real-Time LiDAR Bird's-Eye View Multi-Task Perception Network for Autonomous Driving[J]. 2023. DOI:10.1109 / TITS. 2024. 3510642.) combines geometric features to improve accuracy, while MotionBEV (Zhou B, Xie J, Pan Y, et al. MotionBEV: Attention-aware online LiDAR moving object segmentation with bird's eye view based appearance and motion features[J]. IEEE Robotics and Automation Letters, 2023, 8(12): 8074-8081.) optimizes target recognition by fusing appearance and motion features. Although the BEV method has advantages in computational efficiency and makes up for some of the shortcomings of the RV method, its projection process inevitably introduces information loss, resulting in strong discreteness and directionality in the feature map. Therefore, existing methods do not pay enough attention to these characteristics, which affects the effective modeling of small targets and directional information.

[0004] In the BEV projection process, the extraction and mapping of point cloud features are crucial for subsequent segmentation tasks. As a representative method in the field of point cloud processing, PointNet employs a multilayer perceptron architecture to extract and process point-by-point features from point cloud data, maintaining strong robustness even with data loss as high as 80%. Specifically, in the projection-based online LiDAR-MOS task, its depth needs to be reduced to meet the requirements of the online task; however, reducing the number of MLP layers inevitably leads to performance loss.

[0005] In summary, under the premise of meeting the online segmentation task, existing methods still have low accuracy in segmenting moving targets due to the strong discreteness and directionality of feature maps and the performance loss caused by reducing the number of MLP layers. Therefore, it is necessary to propose a new online segmentation method for moving targets to improve the accuracy of target segmentation. Summary of the Invention

[0006] The purpose of this invention is to address the low accuracy of existing online moving target segmentation methods by proposing an online LiDAR moving target segmentation method based on Haar wavelet enhancement.

[0007] The technical solution adopted by this invention to solve the above-mentioned technical problems is: an online LiDAR moving target segmentation method based on Haar wavelet enhancement, the method specifically includes the following steps:

[0008] Step 1: Use the current frame point cloud data collected by the lidar as the input to the projection network, and output the BEV feature map through the projection network;

[0009] Step 2: Extract the target's motion features based on the current frame point cloud data and the previous frame point cloud data collected by the lidar;

[0010] Step 3: Input the BEV feature map and the target's motion features into the improved Unet bird's-eye view segmentation model, which includes a wavelet fusion block;

[0011] The target segmentation results are output using the improved Unet bird's-eye view segmentation model.

[0012] The beneficial effects of this invention are:

[0013] (1) Optimize point cloud feature extraction. A residual multilayer perceptron structure is adopted. By introducing a residual mechanism, the residual multilayer perceptron structure can reduce the information loss caused by depth reduction, improve the mapping quality from point cloud to BEV feature map, and enhance the expression of spatial information.

[0014] (2) Enhance the directional representation of BEV features. The designed wavelet fusion block can enhance multi-scale feature modeling based on Haar wavelet transform, strengthen the information representation of BEV feature maps in the horizontal, vertical and diagonal directions, and improve the ability to capture moving targets.

[0015] (3) Optimize the segmentation network and improve generalization ability. The improved model proposed in this invention enhances the fusion of motion features and appearance features based on wavelet transform. The optimized encoder-decoder architecture can improve segmentation accuracy while reducing computational overhead, meeting the real-time requirements of online tasks.

[0016] In summary, this invention improves the segmentation accuracy and real-time performance from the perspective of BEV by point cloud feature extraction, feature fusion modeling, and the design of an efficient segmentation network. Attached Figure Description

[0017] Figure 1 This is a flowchart of an online LiDAR moving target segmentation method based on Haar wavelet enhancement according to the present invention;

[0018] In the figure, Q1 represents the feature corresponding to the point cloud in the previous frame, and Q2 represents the feature corresponding to the point cloud in the next frame.

[0019] Figure 2 This is a schematic diagram of the working process of wavelet fusion block (WFB);

[0020] In the figure, WTN represents the wavelet transform unit;

[0021] Figure 3 This is a graph showing the relationship between the number of wavelet fusion blocks, segmentation accuracy, and computational cost. Detailed Implementation

[0022] Specific implementation method one: Combining Figure 1 This embodiment describes an online LiDAR moving target segmentation method based on Haar wavelet enhancement. The method specifically includes the following steps:

[0023] Step 1: Use the current frame point cloud data collected by the lidar as the input to the projection network, and output the BEV feature map through the projection network;

[0024] Step 2: Extract the target's motion features based on the current frame point cloud data and the previous frame point cloud data collected by the lidar;

[0025] Step 3: Input the BEV feature map and the target's motion features into the improved Unet bird's-eye view segmentation model (WT-Unet), which includes wavelet fusion blocks;

[0026] The target segmentation results are output using the improved Unet bird's-eye view segmentation model.

[0027] Specific Implementation Method Two: This implementation method is a further limitation of Specific Implementation Method One. The projection network includes an input layer, a BEV feature mapping layer, and a convergence layer.

[0028] The other steps and parameters are the same as in Specific Implementation Method 1.

[0029] Specific Implementation Method Three: This implementation method further defines Specific Implementation Method Two. The working process of the projection network is as follows:

[0030] The input layer of the projection network is a residual multilayer perceptron, which outputs the features of each point in the current frame point cloud data.

[0031] The BEV feature mapping layer projects each point in the current frame's point cloud data from three-dimensional space to a bird's-eye view grid, forming an efficient two-dimensional feature representation;

[0032] The aggregation layer aggregates the point cloud features within each grid using max pooling. In other words, max pooling preserves the most salient features at each grid location, enhancing feature representation. The aggregation result is then used as the output BEV feature map.

[0033] The other steps and parameters are the same as in Specific Implementation Method Two.

[0034] Specific Implementation Method Four: This implementation method further defines Specific Implementation Method Three. The input of the residual multilayer perceptron is... ,in, This indicates the coordinates of the points collected by the lidar. Indicates coordinates as The reflection intensity value of the point, Indicates coordinates as The depth information of the points.

[0035] The other steps and parameters are the same as in Specific Implementation Method 3.

[0036] Each point acquired by the LiDAR can be converted into an input form for a residual multilayer perceptron (LMP), allowing the LMP to extract features from each point. Compared to traditional LMPs, the residual LMP reduces information loss due to depth reduction by introducing a residual mechanism, improving the mapping quality from point cloud to BEV feature map and enhancing spatial information representation.

[0037] Specific Implementation Method Five: This implementation method is a further limitation of Specific Implementation Method Four, wherein the coordinates are... The depth information of the point is:

[0038]

[0039] The other steps and parameters are the same as in Specific Implementation Method Four.

[0040] Specific Implementation Method Six: This implementation method is a further limitation of Specific Implementation Method Five. The specific process of step two is as follows:

[0041] The previous frame of point cloud data acquired by the lidar is input into the projection network to obtain the BEV feature map corresponding to the previous frame of point cloud data. Then, a residual image is obtained based on the BEV feature map corresponding to the previous frame of point cloud data and the BEV feature map corresponding to the current frame of point cloud data. The residual image is then passed through a first convolutional layer with a kernel size of 3×3 and a first ReLU activation function layer in sequence. The motion features of the target are output through the first ReLU activation function layer.

[0042] The other steps and parameters are the same as in Specific Implementation Method 5.

[0043] Integrating the extracted target motion features into the target segmentation process of the improved Unet bird's-eye view segmentation model can effectively improve the accuracy of target segmentation.

[0044] Specific implementation method seven: Combination Figure 2 This embodiment describes a further limitation of specific embodiment six. The wavelet fusion block includes a first normalization layer, a wavelet transform unit, a second normalization layer, and a channel attention unit (ECA), wherein:

[0045] The input of the wavelet fusion block is used as the input of the first normalization layer (the normalization layer used in this invention is LayerNormalization, LN), and the output of the first normalization layer is used as the input of the wavelet transform unit. The input of the first normalization layer and the output of the wavelet transform unit are residually connected, and the residual connection result is used as the input of the second normalization layer.

[0046] The output of the second normalization layer is used as the input of the channel attention unit. The output of the channel attention unit is then residually concatenated with the input of the second normalization layer. The result of the residual concatenation is then used as the output of the wavelet fusion block.

[0047] The other steps and parameters are the same as in Specific Implementation Method Six.

[0048] The wavelet fusion block designed in this embodiment consists of a wavelet transform unit module and a channel attention module. The wavelet transform unit can extract multi-scale directional features, and the channel attention module can adaptively adjust the channel weights based on the dependencies between channels. Finally, the output features of the channel attention module are residually fused with the original input, so that the features have both directional information and channel optimization capabilities, thereby optimizing the feature representation ability and improving the network's ability to perceive moving targets. At the same time, it optimizes the interaction of channel information and avoids introducing additional parameters and high computational overhead.

[0049] Specific implementation method eight: Combination Figure 2 This embodiment describes a further limitation of specific embodiment seven. The working process of the wavelet transform unit is as follows:

[0050] Within the wavelet transform unit, the input feature map first passes through a wavelet transform layer, which decomposes the input feature map into low-frequency components. and high frequency components , , ;

[0051] Among them, low-frequency components Represents global structural information, high-frequency components Represents horizontal details, high-frequency components Representing vertical details, high-frequency components Indicates details in the diagonal direction;

[0052] Low frequency components The convolutional layer passes through a second convolutional layer with a kernel size of 1×1, a first separable convolutional layer with a kernel size of 3×3, and a second ReLU activation function layer in sequence.

[0053] High frequency components The convolutional layer sequentially passes through a third convolutional layer with a kernel size of 1×1, a second separable convolutional layer with a kernel size of 5×5, and a third ReLU activation function layer.

[0054] High frequency components It sequentially passes through a fourth convolutional layer with a kernel size of 1×1, a third separable convolutional layer with a kernel size of 5×5, and a fourth ReLU activation function layer;

[0055] High frequency components It sequentially passes through a fifth convolutional layer with a kernel size of 1×1, a fourth separable convolutional layer with a kernel size of 5×5, and a fifth ReLU activation function layer;

[0056] Then multiply the outputs of the fourth ReLU activation function layer and the fifth ReLU activation function layer, and pass the multiplication result through the Softmax activation function layer;

[0057] The output of the Softmax activation function layer is multiplied by the output of the third ReLU activation function layer. The result of the multiplication is added to the output of the second ReLU activation function layer. The result of the addition is then used as the input of the inverse wavelet transform layer, and the output of the inverse wavelet transform layer is used as the output of the wavelet transform unit.

[0058] The other steps and parameters are the same as in Specific Implementation Method Seven.

[0059] In this embodiment, after the high-frequency and low-frequency components, a convolutional layer is used to achieve channel fusion, compress redundant information, and improve computational efficiency. A depthwise separable convolutional layer can capture local contextual information (where a 5×5 kernel-size depthwise separable convolutional layer can extract directional information within a larger receptive field), enhancing the ability to express global features. The nonlinear activation function ReLU can further enhance nonlinear expressive power. The vertical component output of the fourth ReLU activation function layer is processed through directional interaction operations. and the horizontal component of the output of the fifth ReLU activation function layer Performing element-wise multiplication can balance high-frequency information in different directions.

[0060] The perfect reconstructibility of Haar wavelet decomposition avoids error accumulation caused by multiple transformations. This characteristic provides theoretical support for feature decomposition and fusion operations, enabling the low-frequency components after decomposition to focus on global structural information, while the high-frequency components focus on capturing directional details, thereby effectively improving the quality of feature representation.

[0061] Specific Implementation Method Nine: This implementation method further defines Specific Implementation Method Eight. The wavelet transform layer performs Haar wavelet decomposition using one low-pass filter and three high-pass filters; the low-pass filter is:

[0062]

[0063] in, Indicates a low-pass filter;

[0064] The high-pass filters are as follows:

[0065]

[0066]

[0067]

[0068] in, , and This indicates a high-pass filter.

[0069] The other steps and parameters are the same as in Specific Implementation Method 8.

[0070] The wavelet transform layer of this invention uses the Haar wavelet transform, based on the low-pass filter. Low-frequency components can be obtained A high-pass filter is used. High-frequency components can be obtained A high-pass filter is used. High-frequency components can be obtained A high-pass filter is used. High-frequency components can be obtained .

[0071] Specific Implementation Method Ten: This implementation method further defines Specific Implementation Method Nine. The working process of the improved Unet bird's-eye view segmentation model is as follows:

[0072] The BEV feature map is passed sequentially through a sixth convolutional layer with a kernel size of 3×3 and a sixth ReLU activation function layer.

[0073] The outputs of the sixth ReLU activation function layer and the first ReLU activation function layer are concatenated to obtain the concatenation result a. The concatenation result a is then passed through the seventh convolutional layer with a kernel size of 3×3 and the seventh ReLU activation function layer in sequence.

[0074] The output of the first ReLU activation function layer is passed sequentially through the eighth convolutional layer with a kernel size of 3×3 and the eighth ReLU activation function layer.

[0075] The output of the seventh ReLU activation function layer is concatenated with the output of the eighth ReLU activation function layer to obtain the concatenated result b. The concatenated result b is then passed through the ninth convolutional layer with a kernel size of 3×3 and the ninth ReLU activation function layer in sequence.

[0076] The output of the eighth ReLU activation function layer is passed sequentially through the tenth convolutional layer with a kernel size of 3×3 and the tenth ReLU activation function layer.

[0077] The output of the ninth ReLU activation function layer is concatenated with the output of the tenth ReLU activation function layer to obtain the concatenation result c. The concatenation result c is then passed through the first wavelet fusion block, the second wavelet fusion block, and the third wavelet fusion block in sequence.

[0078] The output of the third wavelet fusion block is passed sequentially through the eleventh convolutional layer with a kernel size of 3×3 and the eleventh ReLU activation function layer.

[0079] The output of the eleventh ReLU activation function layer is then passed through the first upsampling layer, and the output of the first upsampling layer is then concatenated with the concatenation result c to obtain the concatenation result d; the concatenation result d is then passed through the twelfth convolutional layer with a kernel size of 3×3 and the twelfth ReLU activation function layer in sequence.

[0080] The output of the twelfth ReLU activation function layer is then passed through the second upsampling layer, and the output of the second upsampling layer is then concatenated with the concatenation result b to obtain the concatenation result f; the concatenation result f is then passed through the thirteenth convolutional layer with a kernel size of 3×3 and the thirteenth ReLU activation function layer in sequence.

[0081] The output of the thirteenth ReLU activation function layer is then passed through the third upsampling layer, and the output of the third upsampling layer is then concatenated with the concatenation result a to obtain the concatenation result g. The concatenation result g is then passed through the fourteenth convolutional layer with a kernel size of 3×3 and the fourteenth ReLU activation function layer in sequence.

[0082] The output of the fourteenth ReLU activation function layer is passed through the fourth upsampling layer, and then the output of the fourth upsampling layer is passed through the fifteenth convolutional layer with a kernel size of 3×3 and the fifteenth ReLU activation function layer in sequence.

[0083] The output of the fifteenth ReLU activation function layer is passed through a sixteenth convolutional layer with a kernel size of 1×1, and the output of the sixteenth convolutional layer is used as the output of the improved Unet bird's-eye view segmentation model.

[0084] The other steps and parameters are the same as in Specific Implementation Method Nine.

[0085] The improved Unet bird's-eye view segmentation model used in this invention consists of an encoder, a decoder, and jump connections, with the core module being a wavelet fusion block. Haar wavelet transform is used to capture global structure and local details, improving computational efficiency and feature representation capabilities. The wavelet fusion block can effectively reduce the impact of data discreteness and projection noise.

[0086] (1) Encoder design

[0087] The encoder is responsible for extracting preliminary features from the input BEV feature map and performing downsampling operations. The size of the input feature map is progressively reduced to 1 / 2, 1 / 4, and 1 / 8. The feature map then undergoes wavelet fusion blocks for wavelet space feature fusion. Through wavelet transform, the feature map can capture both global structure and local details. In each wavelet fusion block, the feature map undergoes normalization to reduce feature distribution bias, and then is fused by wavelet transform units.

[0088] (2) Decoder design

[0089] The decoder aims to recover the spatial information lost during the encoder's downsampling process. An upsampling module is used to progressively restore spatial resolution. During each decoding layer, the feature map not only undergoes upsampling but is also fused with features from the corresponding layer in the encoder via skip connections. The decoder can obtain rich low-level feature information from the encoder, which helps recover details and mitigates the vanishing gradient problem that may occur in deep networks. The final layer of the decoder maps the recovered feature map to the final output space through a convolutional operation. Through multi-layered decoding and fusion operations, the spatial detail information of the input image can be efficiently recovered, ultimately achieving bird's-eye view resolution and moving object segmentation.

[0090] (3) Upsampling module

[0091] To achieve efficient reconstruction of image spatial resolution and generate accurate segmentation results, this invention introduces a Cascaded Upsampler (CUP). Each upsampling operation doubles the spatial resolution, extracting features using a 3×3 convolutional layer and adding nonlinear representation using the ReLU activation function. Through multiple upsampling steps, hidden features are restored from low resolution to the target high resolution, gradually restoring the spatial details of the image while preserving semantic information.

[0092] (4) Skip connections and point convolutions

[0093] To compensate for the spatial information lost during downsampling, this invention employs skip connections to enhance feature supervision and adjusts the number of channels through the final convolution. Concatenating the features output from corresponding layers of the encoder and decoder through channel concatenation preserves more detailed information. Assuming the encoder's... Layer feature map is The feature map of the corresponding layer of the decoder is Then the skip connection operation will be executed: Then, the number of channels is adjusted through the last convolutional layer to reduce computation and enhance feature adaptability.

[0094] Model training configuration:

[0095] The proposed model was trained from scratch on the SemanticKITTI dataset. The optimizer employed stochastic gradient descent with a momentum parameter set to 0.9 to accelerate model convergence and improve the stability of gradient updates. To suppress overfitting, the optimizer's weight decay coefficient was set to 0.0001, and the batch size was 8 to ensure stable training under computationally limited conditions. The initial learning rate was set to 0.01, and was dynamically adjusted during training to meet the optimization needs of each training stage. The entire training process consisted of 100 rounds.

[0096] Module effectiveness study:

[0097] To further validate the effectiveness of each module, ablation experiments were conducted to analyze the impact of each module on the overall performance. The experiments primarily evaluated the point cloud feature extraction of the residual multilayer perceptron, the wavelet fusion block, and the improved WT-UNet structure to analyze the contribution of each component to improving segmentation accuracy and computational efficiency. All experiments were performed on the SemanticKITTI validation set, using the average intersection-over-union ratio (IoU) as the core evaluation metric. Inference time was also recorded to measure computational overhead, thus comprehensively analyzing the impact of different modules on model performance.

[0098] (1) Overall model validity analysis

[0099] The overall ablation experiment results are shown in Table 1. The baseline model achieved an mIoU of 76.5% and an inference time of 28.5 ms on the test set. When the corresponding PointNet module was removed from the baseline model and a residual multilayer perceptron was introduced, the mIoU improved to 77.4%, and the inference time decreased to 26.3 ms, indicating that the residual multilayer perceptron can effectively enhance local feature extraction capabilities while optimizing computational efficiency. After replacing the original Unet module with WT-Unet of this invention, the mIoU improved to 79.56%, and the inference time slightly increased to 31.1 ms, indicating that this structure can maintain good computational efficiency while improving feature representation capabilities. Finally, the improved complete model of this invention, which combines the residual multilayer perceptron and WT-UNet, further improved the mIoU to 81.3% and controlled the inference time at 27.1 ms. Compared with the baseline model, it improved accuracy by 4.8% while only increasing computational overhead by 2.1 ms, demonstrating a better balance between efficiency and performance.

[0100] Table 1. Overall Module Effectiveness Study

[0101]

[0102] (2) Wavelet fusion block effectiveness analysis

[0103] To deeply analyze the impact of wavelet fusion blocks and their core components (wavelet transform units) on the overall model performance, the number of wavelet fusion blocks was adjusted to study their role in improving segmentation accuracy and inference efficiency. Experimental results are as follows: Figure 3 As shown, the impact of the number of different wavelet fusion blocks on segmentation accuracy and computational cost is presented in detail.

[0104] Experimental results show that the introduction of wavelet fusion blocks significantly improves segmentation accuracy. However, as the number of wavelet fusion blocks increases, the accuracy gain gradually saturates, while the computational cost increases accordingly. Compared to the baseline model (76.5% mIoU, 28.5ms), introducing one wavelet fusion block increases the mIoU to 77.6% and the inference time to 26.1ms, indicating that the wavelet fusion block effectively enhances feature fusion capabilities and improves segmentation accuracy. Further increasing to three blocks results in an mIoU of 79.56% and an inference time of 27.1ms. However, as the number of wavelet fusion blocks continues to increase, the accuracy improvement gradually plateaus. Increasing to four and five blocks only improves the mIoU to 79.60% and 79.61%, respectively, but the inference time significantly increases to 29.2ms and 31.31ms, indicating that additional wavelet fusion blocks no longer provide a significant accuracy improvement but instead increase computational cost. Analysis of comprehensive experimental data shows that when the number of wavelet fusion blocks is 3, the mIoU reaches 79.56%, and the inference time remains at 27.1 ms, achieving a good balance between accuracy and computational efficiency. Therefore, the model of this invention ultimately selects 3 wavelet fusion blocks as the optimal configuration, which improves segmentation accuracy while ensuring efficient inference performance, providing a better network module design scheme for moving target segmentation tasks.

[0105] Experimental Results and Comparison

[0106] The SemanticKITTI dataset contains a training set (numbered 00-10) and a validation set (numbered 11-21). During model training, only the training set data was used for optimization; the validation set was not involved. To comprehensively evaluate the performance of the proposed model, moving object segmentation prediction results were submitted on the SemanticKITTI-MOS benchmark platform and compared with current mainstream deep learning-based methods. Specifically, MotionBEV was selected as the baseline model, and mainstream methods under three different data representation methods were compared: methods based on range view projection, methods based on bird's-eye view projection, and methods based on point clouds and voxels. Through the comparison of different data representation methods, the segmentation performance of the proposed model under different perspectives and its adaptability to moving object detection tasks were further analyzed.

[0107] (1) Quantitative comparative analysis

[0108] Quantitative analysis can be used to evaluate the performance of the proposed method compared to mainstream methods. Experiments were conducted based on the SemanticKITTI dataset, and statistical analysis was performed on the mean intersection-union ratio (mIoU) of each method on the validation and test sets. The experimental results are shown in Table 2, covering three methods: point cloud / voxel-based, range view projection, and bird's-eye view projection, comprehensively evaluating their performance in moving target segmentation tasks.

[0109] Table 2 Comparison of inference speed and accuracy with the best models on the SemanticKITTI dataset.

[0110]

[0111] In Table 2, the InsMOS model comes from reference 1 (Wang N, Shi C, Guo R, et al. Insmos: Instance-aware moving object segmentation in lidar data[C] / / 2023 IEEE / RSJInternational Conference on Intelligent Robots and Systems (IROS). IEEE,2023: 7598-7605.), the 4DMOS model comes from reference 2 (Mersch B, Chen X, Vizzo I, et al. Receding moving object segmentation in 3d lidar data using sparse 4dconvolutions[J]. IEEE Robotics and Automation Letters, 2022, 7(3): 7503-7510.), and the MambaMOS model comes from reference 3 (Zeng K, Shi H, Lin J, et al. MambaMOS: LiDAR-based 3D moving object segmentation with motion-aware state space model[C] / / Proceedings of the 32nd ACM International Conference on Multimedia. 2024:1505-1513.), the LMNet model comes from reference 4 (Minemura K, Liau H, Monrroy A, et al. LMNet: Real-time multiclass object detection on CPU using 3D LiDAR[C] / / 2018 3rdAsia-Pacific Conference on Intelligent Robot Systems (ACIRS). IEEE, 2018: 28-34.), and the MotionSeg3D-v1 model comes from reference 5 (Sun J, Dai Y, Zhang X, et al.).Efficientspatial-temporal information fusion for lidar-based 3d moving objectsegmentation[C] / / 2022 IEEE / RSJ International Conference on Intelligent Robotsand Systems (IROS). IEEE, 2022: 11456-11463.), the model LiMoSeg is from reference 6 (Mohapatra S, Yogamani S, Kumar V R, et al.LiDAR-BEVMTN: Real-Time LiDARBird's-Eye View Multi-Task Perception Network for Autonomous Driving[J].2023.DOI:10.1109 / TITS.2024.3510642.), and the model MF-MOS is from reference 7 (Cheng J, Zeng K,Huang Z, et al. Mf-mos: A motion-focused model for moving object segmentation[C] / / 2024 IEEE International Conference on Robotics and Automation (ICRA).IEEE, 2024: 12499-12505.).

[0112] Experimental results show that segmentation methods using different strategies differ in terms of accuracy and inference efficiency. Point cloud-based methods perform better in terms of mIoU; for example, 4DMOS achieves an mIoU of 77.2% on the validation set, but its mIoU drops to 65.2% on the test set, indicating room for improvement in generalization ability. InsMOS achieves an mIoU of 70.6% on the test set, a 5.4% improvement over 4DMOS, but its inference time is relatively long (127ms), affecting real-time performance. Range view-based methods have faster inference speeds; for example, LMNet and RVMOS have inference times of 38ms and 29ms, respectively, demonstrating high computational efficiency. However, their segmentation accuracy is lower, with mIoUs of 62.5% and 74.7% on the test set, respectively, indicating insufficient target discrimination ability in complex scenes. Bird's-eye view-based methods achieve a better balance between accuracy and inference efficiency; for example, MotionBEV achieves an mIoU of 69.7% on the test set with an inference time of only 28.5ms, demonstrating good real-time performance. The mIoU of MF-MOS was improved to 76.7%, but the inference time increased to 96.19 ms, resulting in increased computational complexity. The method of this invention outperforms existing methods in both accuracy and inference efficiency. On the test set, the mIoU reaches 75.6%, a 3.9% improvement over the baseline model, and the inference time is controlled at 27.1 ms, a reduction of 1.4 ms, demonstrating improved computational efficiency. Furthermore, the mIoU on the validation set reaches 81.3%, outperforming all comparative methods, further validating its performance in different scenarios.

[0113] (2) Qualitative analysis

[0114] The performance of the model of this invention in terms of long-distance target detection, target contour integrity and background misclassification is compared with that of LMNet and MotionBEV to evaluate the segmentation ability of the model of this invention in complex scenes.

[0115] Experimental results demonstrate that the model of this invention exhibits significant advantages in occluded target detection, boundary detail processing, and background misclassification. Specifically, LMNet's range-view projection method is susceptible to occlusion interference in complex scenes, leading to numerous misclassifications of targets behind occluded objects, especially in dense traffic environments where foreground targets easily occlude moving targets behind them, thus affecting the overall segmentation performance. In contrast, the model of this invention relies on bird's-eye view projection modeling, effectively avoiding occlusion interference and making the detection of distant targets more stable. Furthermore, compared to MotionBEV, the model of this invention provides more refined boundary detail processing, effectively reducing the blurring and fragmentation of target contours, resulting in more complete segmentation results and improved target shape preservation. Simultaneously, the model of this invention also shows significant improvement over other methods in addressing background misclassification, reducing the probability of static backgrounds being misclassified as moving targets and improving overall recognition accuracy.

[0116] The model of this invention can more effectively separate motion features and appearance features in BEV space, achieving higher segmentation accuracy and exhibiting better adaptability and generalization ability in complex scenarios.

[0117] (3) Runtime efficiency analysis

[0118] To evaluate the computational efficiency of different methods, the average inference time of each model on the SemanticKITTI validation set was measured. Local experiments were conducted on an Intel Core i9-12900K CPU and NVIDIA RTX 3090 GPU. The experimental results show that there are significant differences in inference speed among different methods, and inference overhead directly affects the real-time performance of the model.

[0119] Point cloud-based methods (InsMOS, 4DMOS) have high computational complexity, with inference times of 127ms and 132ms respectively, making them unsuitable for the real-time requirements of automotive scenarios. Range view-based methods offer high inference efficiency; for example, RVMOS has an inference time of only 29ms, but its segmentation accuracy is relatively low, with a test set mIoU of only 74.7%. In contrast, BEV-based methods achieve a better balance between computational efficiency and accuracy. MotionBEV (Baseline) has an inference time of 28.5ms, but its mIoU is only 76.5%. MF-MOS improves accuracy to 76.7%, but its inference time increases to 96.19ms, resulting in significant computational overhead. The model proposed in this invention has an inference time of only 27.1ms, achieving an mIoU of 81.3% while maintaining computational efficiency, thus realizing efficient and accurate moving target segmentation and meeting the needs of online LiDAR motion segmentation tasks.

[0120] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An online LiDAR moving target segmentation method based on Haar wavelet enhancement, characterized in that, The method specifically includes the following steps: Step 1: Use the current frame point cloud data collected by the lidar as the input to the projection network, and output the BEV feature map through the projection network; Step 2: Extract the target's motion features based on the current frame point cloud data and the previous frame point cloud data collected by the lidar; Step 3: Input the BEV feature map and the target's motion features into the improved Unet bird's-eye view segmentation model, which includes a wavelet fusion block; The target segmentation results are output using the improved Unet bird's-eye view segmentation model.

2. The online LiDAR moving target segmentation method based on Haar wavelet enhancement according to claim 1, characterized in that, The projection network includes an input layer, a BEV feature mapping layer, and a convergence layer.

3. The online LiDAR moving target segmentation method based on Haar wavelet enhancement according to claim 2, characterized in that, The working process of the projection network is as follows: The input layer of the projection network is a residual multilayer perceptron, which outputs the features of each point in the current frame point cloud data. The BEV feature mapping layer projects each point in the current frame point cloud data from 3D space to a bird's-eye view grid. The aggregation layer aggregates the point cloud features within each grid using max pooling, and the aggregation result is used as the output BEV feature map.

4. The online LiDAR moving target segmentation method based on Haar wavelet enhancement according to claim 3, characterized in that, The input to the residual multilayer perceptron is ,in, This indicates the coordinates of the points collected by the lidar. Indicates coordinates as The reflection intensity value of the point, Indicates coordinates as The depth information of the points.

5. The online LiDAR moving target segmentation method based on Haar wavelet enhancement according to claim 4, characterized in that, The coordinates are The depth information of the point is: 。 6. The online LiDAR moving target segmentation method based on Haar wavelet enhancement according to claim 5, characterized in that, The specific process of step two is as follows: The previous frame of point cloud data acquired by the lidar is input into the projection network to obtain the BEV feature map corresponding to the previous frame of point cloud data. Then, a residual image is obtained based on the BEV feature map corresponding to the previous frame of point cloud data and the BEV feature map corresponding to the current frame of point cloud data. The residual image is then passed through a first convolutional layer with a kernel size of 3×3 and a first ReLU activation function layer in sequence. The motion features of the target are output through the first ReLU activation function layer.

7. The online LiDAR moving target segmentation method based on Haar wavelet enhancement according to claim 6, characterized in that, The wavelet fusion block includes a first normalization layer, a wavelet transform unit, a second normalization layer, and a channel attention unit, wherein: The input of the wavelet fusion block is used as the input of the first normalization layer, and the output of the first normalization layer is used as the input of the wavelet transform unit. The input of the first normalization layer and the output of the wavelet transform unit are residually connected, and the residual connection result is used as the input of the second normalization layer. The output of the second normalization layer is used as the input of the channel attention unit. The output of the channel attention unit is then residually concatenated with the input of the second normalization layer. The result of the residual concatenation is then used as the output of the wavelet fusion block.

8. The online LiDAR moving target segmentation method based on Haar wavelet enhancement according to claim 7, characterized in that, The working process of the wavelet transform unit is as follows: Within the wavelet transform unit, the input feature map first passes through a wavelet transform layer, which decomposes the input feature map into low-frequency components. and high frequency components , , ; Low frequency components The convolutional layer passes through a second convolutional layer with a kernel size of 1×1, a first separable convolutional layer with a kernel size of 3×3, and a second ReLU activation function layer in sequence. High frequency components The convolutional layer sequentially passes through a third convolutional layer with a kernel size of 1×1, a second separable convolutional layer with a kernel size of 5×5, and a third ReLU activation function layer. High frequency components It sequentially passes through a fourth convolutional layer with a kernel size of 1×1, a third separable convolutional layer with a kernel size of 5×5, and a fourth ReLU activation function layer; High frequency components It sequentially passes through a fifth convolutional layer with a kernel size of 1×1, a fourth separable convolutional layer with a kernel size of 5×5, and a fifth ReLU activation function layer; Then multiply the outputs of the fourth ReLU activation function layer and the fifth ReLU activation function layer, and pass the multiplication result through the Softmax activation function layer; The output of the Softmax activation function layer is multiplied by the output of the third ReLU activation function layer. The result of the multiplication is added to the output of the second ReLU activation function layer. The result of the addition is then used as the input of the inverse wavelet transform layer, and the output of the inverse wavelet transform layer is used as the output of the wavelet transform unit.

9. The online LiDAR moving target segmentation method based on Haar wavelet enhancement according to claim 8, characterized in that, The wavelet transform layer performs Haar wavelet decomposition using one low-pass filter and three high-pass filters; the low-pass filter is: in, Indicates a low-pass filter; The high-pass filters are as follows: in, , and This indicates a high-pass filter.

10. The online LiDAR moving target segmentation method based on Haar wavelet enhancement according to claim 9, characterized in that, The improved Unet bird's-eye view segmentation model works as follows: The BEV feature map is passed sequentially through a sixth convolutional layer with a kernel size of 3×3 and a sixth ReLU activation function layer. The outputs of the sixth ReLU activation function layer and the first ReLU activation function layer are concatenated to obtain the concatenation result a. The concatenation result a is then passed through the seventh convolutional layer with a kernel size of 3×3 and the seventh ReLU activation function layer in sequence. The output of the first ReLU activation function layer is passed sequentially through the eighth convolutional layer with a kernel size of 3×3 and the eighth ReLU activation function layer. The output of the seventh ReLU activation function layer is concatenated with the output of the eighth ReLU activation function layer to obtain the concatenated result b. The concatenated result b is then passed through the ninth convolutional layer with a kernel size of 3×3 and the ninth ReLU activation function layer in sequence. The output of the eighth ReLU activation function layer is passed sequentially through the tenth convolutional layer with a kernel size of 3×3 and the tenth ReLU activation function layer. The output of the ninth ReLU activation function layer is concatenated with the output of the tenth ReLU activation function layer to obtain the concatenation result c. The concatenation result c is then passed through the first wavelet fusion block, the second wavelet fusion block, and the third wavelet fusion block in sequence. The output of the third wavelet fusion block is passed sequentially through the eleventh convolutional layer with a kernel size of 3×3 and the eleventh ReLU activation function layer. The output of the eleventh ReLU activation function layer is then passed through the first upsampling layer, and the output of the first upsampling layer is then concatenated with the concatenation result c to obtain the concatenation result d; the concatenation result d is then passed through the twelfth convolutional layer with a kernel size of 3×3 and the twelfth ReLU activation function layer in sequence. The output of the twelfth ReLU activation function layer is then passed through the second upsampling layer, and the output of the second upsampling layer is then concatenated with the concatenation result b to obtain the concatenation result f; the concatenation result f is then passed through the thirteenth convolutional layer with a kernel size of 3×3 and the thirteenth ReLU activation function layer in sequence. The output of the thirteenth ReLU activation function layer is then passed through the third upsampling layer, and the output of the third upsampling layer is then concatenated with the concatenation result a to obtain the concatenation result g. The concatenation result g is then passed through the fourteenth convolutional layer with a kernel size of 3×3 and the fourteenth ReLU activation function layer in sequence. The output of the fourteenth ReLU activation function layer is passed through the fourth upsampling layer, and then the output of the fourth upsampling layer is passed through the fifteenth convolutional layer with a kernel size of 3×3 and the fifteenth ReLU activation function layer in sequence. The output of the fifteenth ReLU activation function layer is passed through a sixteenth convolutional layer with a kernel size of 1×1, and the output of the sixteenth convolutional layer is used as the output of the improved Unet bird's-eye view segmentation model.

Images