Monocular BEV perception lightweight method for automatic driving low-resource platform

Abstract

The application discloses a monocular BEV perception lightweight method for an automatic driving low-resource platform, and specifically comprises the following steps: acquiring an image data set, pre-processing the data set, and training a high-performance teacher network; through a four-fold collaborative compression strategy of backbone lightweight, channel pruning, spatial resolution optimization, and interpolation operator replacement, a student network is constructed, and steps of image feature extraction, depth estimation, BEV space projection, BEV feature coding, and multi-task output are sequentially executed, and knowledge distillation is performed on the student network based on the teacher network; after the student network is trained, baseline performance testing and hyperparameter tuning are performed to evaluate and optimize the model performance, and blind test submission and result analysis are used to verify the scene understanding performance of the student network, so that lightweight and efficient perception are realized under the premise that the precision remains unchanged.

Description

Technical Field

[0001] This application relates to the technical field of reducing model computation overhead while ensuring detection accuracy in autonomous driving, specifically to a lightweight monocular BEV perception method for low-resource autonomous driving platforms. Background Technology

[0002] In autonomous driving environments, targets are diverse, including cars, pedestrians, and motorcycles, with significant differences in scale and appearance. Simultaneously, the background environment of the targets to be detected is complex and ever-changing, necessitating both faster detection speeds and high accuracy. However, the widespread adoption of advanced intelligent driving functions is currently severely limited by the high cost of perception systems, making pure vision-based solutions a more feasible option.

[0003] The release of the NuScenes dataset provided a standardized benchmark for monocular 3D perception and spurred the development of LSS (Local Sequence Segmentation). LSS was the first to demonstrate that effective bird's-eye view representations and 3D object detection could be generated using only monocular images through implicit depth estimation and geometric projection. Subsequently, BEVDepth improved monocular BEV detection performance by optimizing depth supervision strategies; MonoBEV introduced multi-view geometric constraints to further alleviate the monocular depth blur problem; BEVDistill introduced distillation into the BEV domain, but it focused on multi-camera cross-modal distillation, did not consider monocular scenes, and the model still relied on a complex Transformer architecture, resulting in high computational overhead. How to handle monocular, LiDAR-free environments remains an unexplored area. However, with breakthroughs in model compression technology, knowledge distillation has demonstrated powerful capabilities in image classification, object detection, and semantic segmentation since its inception. Its core idea is to use a "teacher-student learning" mechanism, with LSS or BEVDepth as the teacher network model, to design a lightweight student network based on MobileNetV3 or EfficientNet-Lite, and to introduce multi-level distillation constraints in the BEV feature space, such as feature alignment, attention transfer, and output calibration, so that the student network model can not only imitate the detection results, but also learn the teacher network model's understanding of 3D spatial structure.

[0004] While existing technologies have made significant progress in BEV 3D target detection, such as BEVFormer and UniAD models which achieve high-precision perception through multi-camera and temporal modeling, their high computational complexity and reliance on multi-sensor input make them difficult to adapt to automotive embedded platforms. Existing lightweight solutions, while reducing the number of parameters, generally lack knowledge distillation mechanisms for monocular, non-temporal scenarios. Their compression strategies are fragmented and lack systematic coordination, and they do not optimize key operators such as interpolation and convolution for edge hardware, resulting in significant accuracy loss and low deployment efficiency. Therefore, these methods are not ideal. The method introduced in this invention, however, achieves good speed while maintaining relatively good accuracy. Summary of the Invention

[0005] To address the aforementioned issues, this invention primarily aims to resolve some minor problems in monocular BEV perception under autonomous driving scenarios. For example, in low-computing-power environments, hardware limitations affect the speed and accuracy of monocular BEV perception. Therefore, based on the core ideas of the LSS framework, a lightweight student network architecture is innovatively constructed. Through a four-pronged collaborative strategy of backbone replacement, channel compression, resolution adjustment, and interpolation operator reconstruction, the computational load and memory consumption are systematically reduced. Simultaneously, addressing the issues of weak semantic abstraction capabilities and inaccurate spatial reasoning under monocular input, a knowledge distillation mechanism adapted to the BEV feature space is designed, encompassing both feature layer alignment and output layer distribution transfer paths. This achieves effective transfer of semantic discrimination knowledge while freezing the teacher network model. Finally, the lightweight structural design and distillation training mechanism are deeply integrated to form a lightweight monocular BEV perception method. The specific steps are as follows: An image dataset is acquired, preprocessed, and a high-performance teacher network is trained. A student network is constructed using a four-fold collaborative compression strategy of backbone lightweighting, channel pruning, spatial resolution optimization, and interpolation operator replacement. The student network is then sequentially executed with image feature extraction, depth estimation, BEV spatial projection, BEV feature encoding, and multi-task output steps. Knowledge distillation is then performed on the student network based on the teacher network. After the student network is trained, the model performance is evaluated and optimized through baseline performance testing and hyperparameter tuning. Blind test submission and result analysis are used to verify the scene understanding performance of the student network, achieving lightweight and efficient perception while maintaining the same accuracy.

[0006] Furthermore, the image feature extraction specifically includes: MobileNetV3-Large is used as the backbone network, and lightweight feature extraction is achieved through an inverted residual bottleneck structure, SE module, and Hard-Swish activation. A lightweight feature pyramid network containing only three layers (P3-P5) is then constructed thereafter. 1. Convolution for channel alignment, 3. 3. Depthwise separable convolutions are used for feature fusion, and multi-scale features are uniformly upsampled to the original value. Figure 1 / 8 resolution.

[0007] Furthermore, the depth estimation specifically includes: The number of depth prediction channels was reduced from 118 to 32 uniform depth intervals; the standard 3 in the depth branch was also reduced. Replace 3 convolutions with depthwise convolutions plus 1 A combination structure of 1-point convolution is used; the depth probability is normalized by Softmax along the depth dimension; the pixel features are copied into the 3D space voxel by weighting according to the depth probability based on the camera intrinsic and extrinsic parameters, realizing the geometric dimensionality upgrade from 2D to 3D.

[0008] Furthermore, the BEV spatial projection specifically includes: Three-dimensional voxels are average-pooled along the height dimension to generate two-dimensional BEV feature maps, with the resolution set to 100. 100; Nearest neighbor interpolation is used instead of bilinear interpolation in voxel-to-raster mapping; After projection, Layer Normalization is performed on the BEV features to normalize the channels.

[0009] Furthermore, the BEV feature encoding step includes: Use 1-2 layers of 3 A lightweight convolutional structure consisting of 3 convolutions, BatchNorm, and ReLU is used to encode BEV features. The number of channels is fixed at 128, and no Transformer, Deformable Convolution, or positional encoding mechanisms are introduced.

[0010] Furthermore, the multi-task output step specifically includes: Based on the same set of encoded BEV features, respectively through 1 1. Convolutional methods are used to construct a 3D object detection head and a semantic map segmentation head. The semantic map segmentation head is a lightweight FCN structure that outputs probability maps of semantic categories such as lane lines, sidewalks, drivable areas, and parking lines.

[0011] Furthermore, the knowledge distillation specifically includes: constructing a two-path distillation framework for teacher network models and student network models, including: Feature distillation: Aligning intermediate representations of teachers and students in the BEV feature space; Output distillation: Align the predicted distributions of teachers and students at the task output layer to take advantage of the class relationships and uncertainty information implied by the soft labels in the teacher output.

[0012] Furthermore, the feature distillation specifically involves extracting BEV feature maps of both teachers and students after the BEV encoder output, with the feature map size being H. W C, calculate the L2 distance loss pixel by pixel, the specific formula is: ; in, The feature distillation loss is represented by H, where H represents the height of the BEV feature map, W represents the width of the BEV feature map, C represents the number of channels in the BEV feature map, and i represents the index in the height direction of the feature map. i H,j represents the index along the width direction of the feature map, 1 j W, k represents the index of the feature map channel direction, 1 k C, This represents the feature value of the student network model at position (i,j,k). This represents the feature value of the teacher network model at position (i,j,k).

[0013] Furthermore, the output distillation uses temperature-scaled KL divergence loss for the heatmap and BEV segmentation probability map in 3D detection, with the specific formula as follows: ; Where, p( =softmax(Y / ), representing the soft probability distribution after temperature scaling, and Y representing the original values ​​of the model. This represents a temperature parameter, a scalar value greater than 0, and 'c' represents the category index. This represents the probability output of the teacher network model in class c. This represents the probability output of the student network model in class c. This represents the output distillation loss, and KL divergence is used to measure the difference between the output probability distributions of teachers and students.

[0014] According to a second aspect of the present invention, a computer program product is provided, on which one or more computer programs are stored, which, when executed by a computer processor, implement the method described above.

[0015] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: In the industry chain, this invention adopts a modular, standardized, and interface-based architecture design, possessing strong horizontal scalability and vertical evolution capabilities: the BEV feature map output dimension is fixed, naturally supporting late fusion with millimeter-wave radar BEV features, realizing a progressive upgrade path of "vision + radar". The semantic map segmentation head output structure is compatible with high-precision map standards such as OpenDRIVE and Lanelet2, and can be directly connected to online mapping or local path planning modules. The 3D detection head branch design follows the mainstream industry paradigm FCOS3D, facilitating seamless integration with downstream tracking, prediction, and decision-making modules. The knowledge distillation framework supports a cloud-edge collaborative paradigm of "pre-training of the teacher network model in the cloud and fine-tuning and deployment of the student network model at the edge," aligning with the new electronic and electrical architecture trend of "central brain + regional terminal" in automakers. Moreover, it can not only be delivered as an independent perception module, but also serve as a "standardized perception base" for intelligent driving software platforms, allowing for customized tailoring and secondary development by different hardware platforms. Attached Figure Description

[0016] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Other embodiments and many anticipated advantages of the embodiments will be readily recognized as they become better understood through reference to the following detailed description. Elements in the drawings are not necessarily to scale. The same reference numerals refer to corresponding similar parts.

[0017] Figure 1 This is a system architecture diagram of a lightweight monocular BEV perception method for low-resource autonomous driving platforms according to an embodiment of this application. Figure 2 This is a student network structure diagram of a lightweight monocular BEV perception method for low-resource autonomous driving platforms according to an embodiment of this application. Figure 3 This is a schematic diagram of the knowledge distillation mechanism of a lightweight monocular BEV perception method for low-resource autonomous driving platforms according to an embodiment of this application. Figure 4 This is a teacher network structure diagram of a lightweight monocular BEV perception method for low-resource autonomous driving platforms according to an embodiment of this application. Figure 5 This is a comparative diagram of four compression strategies for a lightweight monocular BEV perception method for low-resource autonomous driving platforms according to an embodiment of this application. Figure 6 This is a schematic diagram of the structure of a computer system suitable for implementing the electronic device of the present invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] Figure 1 This is a system architecture diagram of a lightweight monocular BEV perception method for low-resource autonomous driving platforms according to an embodiment of this application, as shown below. Figure 1 As shown: By constructing a student network that is attention-free, time-series-free, and independent of multi-sensor dependencies, such as Figure 2 As shown, the input image is acquired from a forward-looking monocular camera and used as the input data source for the student model. The input image is fed into the MobileNetV3-Large lightweight convolutional neural network backbone. Multi-level visual features are extracted using a NAS-optimized inverse residual structure, SE module, and lightweight activation functions ReLU / Hard-Swish to reduce computational overhead while maintaining feature quality. The different levels of features output from the backbone network are then fed into a lightweight feature pyramid network (FPN). Channel alignment is performed using 1×1 convolutions, and feature fusion is achieved using depthwise separable convolutions to obtain multi-resolution semantic features. These fused features serve as input for subsequent depth prediction and geometric projection.

[0021] The depth probability distribution of each pixel is predicted using a depth estimation head: the depth interval is compressed to 32 channels; all convolutional structures use depth-separable convolutions to reduce computational consumption. Then, based on camera calibration parameters, the features of each pixel are projected into a three-dimensional voxel space according to their depth probability, realizing the geometric dimensionality upgrade from two-dimensional image to three-dimensional structure.

[0022] The 3D voxel features are aggregated along the height dimension, and a nearest neighbor interpolation strategy is used to map the voxels onto the BEV plane. The default resolution of the resulting BEV feature map is 100×100 to effectively reduce memory and computational burden. The projected BEV features are fed into a lightweight BEV encoder for spatial semantic encoding: the encoder uses a 1-2 layer inverse residual convolutional structure; it does not use Transformer or computationally expensive convolutions; the aim is to extract local spatial context and maintain low latency.

[0023] Based on shared BEV encoded features, multi-task prediction is performed using a lightweight detection segmentation head: The 3D target detection head outputs information such as the position and size of targets in the BEV space; The number of segmentation head channels is halved, and the output BEV semantic map includes lane lines, road areas, and drivable areas.

[0024] The output results constitute the final perception results of the student model.

[0025] With only a single frame of forward-looking camera image as input, a four-fold collaborative compression strategy—root lightweighting, channel dimension pruning, spatial resolution optimization, and interpolation operator replacement—can systematically reduce computational complexity and memory usage while retaining effective representation capabilities for 3D object detection and semantic map segmentation tasks. It mainly consists of five functional modules: image feature extraction module, depth estimation, BEV spatial projection module, BEV feature encoding module, and multi-task output head module.

[0026] In a specific embodiment, the four-fold collaborative compression strategy is as follows: backbone lightweighting, replacing the traditional ResNet with MobileNetV3-Large; channel dimension pruning, compressing the depth estimation channels from 118 to 32, and halving the number of channels in the multi-task segmentation head; spatial resolution optimization, reducing the default resolution of the BEV feature map from 200×200 to 100×100, and retaining only layers P3 to P5 in the feature pyramid network FPN; and interpolation operator replacement, using nearest neighbor interpolation instead of bilinear interpolation in voxel-to-raster mapping.

[0027] In the image feature extraction stage, this invention uses MobileNetV3-Large as the backbone network to replace the traditional large ResNet structure. It employs a neural architecture search-optimized inverted residual bottleneck block, combined with SE attention and a Hard-Swish activation function. Its high accuracy-low parameter ratio has been verified in the ImageNet classification task. To adapt to the multi-scale semantic information requirements of the BEV perception task, this invention connects a lightweight feature pyramid network afterward, retaining only three layers (P3-P5), using 1... 1. Convolution for channel alignment, 3. 3. Depthwise separable convolutions are used for feature fusion to avoid the additional computational burden brought by standard convolutions; the final output multi-scale feature maps are uniformly upsampled to 1 / 8 of the original image resolution and used as input for the subsequent depth estimation module.

[0028] In the depth estimation stage, the depth prediction branch in the original LSS framework is significantly compressed: firstly, the number of depth distribution channels is reduced from 118 to 32 evenly divided depth intervals, significantly reducing the memory overhead of subsequent dimensionality upscaling operations; secondly, the original standard 3 Replace all 3 convolutions with Depthwise Convolution+1 The PointwiseConvolution architecture, with the former responsible for spatial feature extraction and the latter for inter-channel information fusion, reduces multiplication and addition operations by more than 75% compared to standard convolution. Furthermore, in the depth probability normalization stage, Softmax is used to model the probability distribution along the depth dimension, ensuring that each pixel generates a reasonable depth confidence score. Subsequently, based on the camera's intrinsic and extrinsic parameter matrices, each pixel is copied along its ray direction with depth probability weights to a 3D spatial voxel, completing the geometric dimensionality upgrade from the 2D image plane to 3D physical space.

[0029] During the BEV spatial projection stage, a hardware-friendly modification is performed to the original bilinear interpolation strategy. 3D spatial voxels are compressed using average pooling along the height dimension (Z-axis) to generate a 2D BEV feature map. To adapt to the computational constraints of different deployment platforms, the default value is set to 100. 100, compared to the traditional 200 The 200 configuration reduces the computational load of the subsequent BEV encoder by 75%. Secondly, during projection, computationally intensive bilinear interpolation is abandoned in favor of nearest-neighbor interpolation to map voxels to grids. While sacrificing some spatial smoothness, this significantly improves inference speed and reduces GPU / CPU cache pressure. After projection, the BEV feature maps undergo channel normalization using Layer Normalization to stabilize the training process and accelerate convergence.

[0030] In the BEV feature encoding stage, complex Transformer or Deformable Conv structures are abandoned, and only 1 to 2 layers of standard 3D are retained. The system uses three convolutional layers, each followed by BatchNorm and ReLU activation functions, with a fixed number of 128 channels to avoid high-cost operations such as positional encoding and attention weight calculation. This design aims to preserve local spatial context information while maximizing the compression of model parameters and inference latency. The encoded BEV feature maps are fed into two parallel task heads: a 3D object detection head and a semantic map segmentation head. All branches share the same set of BEV feature inputs, processed through 1... 1. Independent convolutional outputs; the semantic map segmentation head adopts a lightweight FCN structure, outputting probability maps of multiple semantic labels such as lane lines, sidewalks, drivable areas, and parking lines, with resolution comparable to BEV features. Figure 1 To.

[0031] The above-described architecture together constitutes a simplified student network, which significantly improves the computational speed of the model while maintaining essentially the same accuracy.

[0032] To compensate for the loss of representational ability caused by the structural simplification of lightweight student networks, a knowledge distillation framework is constructed for monocular, time-independent, and LiDAR-free input conditions. This framework improves the perceptual performance of student network models with limited capacity by transferring spatial semantic knowledge and output decision knowledge from high-performance teacher network models. Its core includes a "dual-path distillation" strategy: feature distillation, which involves aligning intermediate representations in the BEV feature space, and output distillation, which involves aligning the predicted distribution in the task output layer. This design is motivated by two points: 1) The BEV space is the core hub for monocular-to-3D transformation, and its features contain rich geometric and semantic information; direct alignment can compensate for the insufficient backbone capability of the student network; 2) The soft labels output by the teacher, such as classification confidence, contain inter-class relationships and uncertainties, which are more conducive to small models learning decision boundaries than hard labels. The synergistic effect of the dual paths significantly improves distillation efficiency and final performance.

[0033] Specific implementation of the distillation strategy: Feature distillation: Extract the BEV feature maps of teachers and students after the BEV encoder output. The feature map size is H. W C, calculate the L2 distance loss pixel by pixel, the specific formula is: ; in, The feature distillation loss is represented by H, where H represents the height of the BEV feature map, W represents the width of the BEV feature map, C represents the number of channels in the BEV feature map, and i represents the index in the height direction of the feature map. i H,j represents the index along the width direction of the feature map, 1 j W, k represents the index of the feature map channel direction, 1 k C, This represents the feature value of the student network model at position (i,j,k). This represents the feature value of the teacher network model at position (i,j,k).

[0034] If the student BEV resolution is low, such as below 100×100, then the teacher features are first bilinearly downsampled to align the spatial size.

[0035] The output distillation is applied to the heatmap and BEV segmentation probability map in 3D detection, using temperature-scaled KL divergence loss, with the specific formula as follows: ; Where p( =softmax(Y / ), representing the soft probability distribution after temperature scaling, and Y representing the original values ​​of the model. This represents a temperature parameter, a scalar value greater than 0, and 'c' represents the category index. This represents the probability output of the teacher network model in class c. This represents the probability output of the student network model in class c. This represents the output distillation loss, and KL divergence is used to measure the difference between the output probability distributions of teachers and students.

[0036] During training, the teacher network model is frozen throughout, and only student gradients are backpropagated (default). =1.0, =2.0, emphasizing the importance of output distribution shift; the total loss is a weighted sum of task loss and distillation loss: .

[0037] like Figure 3 As shown, a flowchart of knowledge distillation constructed according to an embodiment of the present invention includes the following specific steps: A high-performance teacher network (in this embodiment, the BEVFormer model with a temporal multi-camera Transformer architecture is used as an example; in other parallel embodiments, a high-resolution LSS model based on a high-capacity ResNet backbone can also be used) accepts multi-camera images and historical frame information as input, and generates BEV feature maps (with a size of 200) through forward propagation. 200) and Heatmap (after Softmax or temperature scaling). The student network model takes a single frame image as input, extracts image features through its backbone and FPN, and generates a BEV feature map (size 100) through the BEV encoder and Depth+Lift-Splat module. 100). To ensure the output characteristics of the teacher and student network models. Figure 1 Therefore, the BEV feature maps of the teacher network model are downsampled to 100 using bilinear interpolation. 100, aligned with the feature map of the student network model, to prepare for subsequent loss calculation.

[0038] Mission loss The calculation is performed based on the 3D detection task of the student network model; then, the feature loss is calculated. The L2 loss is used as the L2 loss between the teacher's and student's BEV feature maps; finally, the output loss is used. The difference between the teacher's and student's heatmap outputs is measured using KL divergence, with a total loss of [missing value]. This loss value guides backpropagation, updating only the parameters of the student network model while keeping the parameters of the teacher network model unchanged. Ultimately, through repeated iterative training, the student network model gradually converges in the 3D detection task, improving its performance.

[0039] This operation allows student network models to approximate the performance of teacher network models without increasing inference burden. This distillation scheme does not rely on complex relation modeling or adversarial training, has low computational cost, and achieves good results.

[0040] After the student network is trained, the model performance is evaluated and optimized through baseline performance testing and hyperparameter tuning. Blind test submission and result analysis are used to verify the scene understanding performance of the student network, achieving lightweight and efficient perception while maintaining the same accuracy.

[0041] Furthermore, after acquiring the dataset, a teacher network model is simultaneously trained, such as... Figure 4 As shown, a high-capacity backbone network and a high-resolution BEV space are employed to provide high-quality spatial semantic knowledge and decision-making knowledge for the lightweight student network model. The training process is as follows: The receiving size is 3 224 A 224-channel image is used as network input, and the input image is fed into a 7-channel... 7. An initial convolutional layer with a stride of 2 is used to extract low-level spatial features and reduce the feature map size to half of its original size. BatchNorm2d batch normalization is applied to the convolutional output to stabilize the feature distribution, followed by a ReLU activation function to enhance the non-linear expressive power of the features. The feature map is then input into a 3-channel input. 3. A max pooling layer with a stride of 2 further compresses the spatial resolution of the feature map, providing a larger receptive field for subsequent deeper network layers.

[0042] The network is divided into four stages according to the residual structure, which progressively deepen and abstract the image features: Stage 1: Contains 3 Bottleneck residual blocks, with an output feature dimension of 256; Stage 2: Contains 4 Bottleneck residual blocks, increasing the output feature dimension to 512; Stage 3: Contains 6 Bottleneck residual blocks to extract higher semantic level features, with an output dimension of 1024; Stage 4: Contains 3 Bottleneck residual blocks, outputting high-dimensional semantic features for the final stage, with the number of channels increased to 2048.

[0043] The above four stages realize multi-layer semantic feature extraction from shallow to deep, enabling the network to capture local texture, structural relationships and global contextual information at the same time.

[0044] The 2048-dimensional feature map output from Stage4 is input into the AdaptiveAvgPool2d layer, and the spatial dimension is compressed into a global feature vector through adaptive average pooling, thereby enhancing the overall scene representation capability.

[0045] A fully connected layer is used to further compress and map the pooled global features to generate a fixed-length high-level representation. The semantic features output by the final fully connected layer are then used as input to the downstream lightweight LSS (Lift-Splat-Shoot) module or other 3D perception task modules.

[0046] like Figure 5 As shown in the diagram, a comparison of four compression strategies based on a specific embodiment of the present invention is presented. Traditional solutions using backbone networks such as ResNet-50 or Swin-T suffer from large parameter counts and computational density, failing to meet the demands of real-time in-vehicle inference. The present invention, however, utilizes MobileNetV3-Large, specifically designed for mobile devices, and leverages depthwise separable convolutions and Neural Architecture Search (NAS) techniques to achieve a better balance between accuracy and efficiency. Furthermore, the 118-channel standard convolution in the depth estimation module is compressed to 32 channels, and Depthwise+Pointwise convolutions are employed, reducing the number of parameters while retaining sufficient feature representation capabilities.

[0047] This invention further improves system performance by optimizing the BEV grid resolution and projection interpolation operator. Traditional solutions use a fixed 200... A 200-resolution BEV raster is computationally intensive and unsuitable for resource-constrained platforms. This invention reduces the BEV raster resolution to 100. The system achieves a accuracy of 100 and supports a dynamic switching mechanism, thereby reducing computational load by approximately 75% while maintaining basic perception accuracy. Regarding BEV projection interpolation, traditional methods use bilinear interpolation, which, while providing good smoothing, is computationally complex and increases cache pressure. This invention employs a nearest-neighbor interpolation method, avoiding multiplication operations and significantly reducing computational complexity and memory overhead. Through these optimizations, this invention constructs a highly efficient BEV perception system with strong prospects for in-vehicle deployment.

[0048] As the global automotive industry accelerates its transformation towards intelligence, electrification, and connectivity, advanced autonomous driving functions are gradually moving from being exclusive to luxury models to the mainstream family car market priced between 100,000 and 200,000 yuan. However, current L2+ / L3 level systems generally rely on sensor stacking solutions such as "multiple cameras + millimeter-wave radar + ultrasonic + high-precision positioning" or even "LiDAR," resulting in high per-vehicle perception BOM costs and severely hindering large-scale adoption. In contrast, the lightweight monocular BEV perception solution proposed in this paper theoretically requires only a single forward-facing camera and a mid-to-low-end domain controller to achieve near-L3 level environmental understanding capabilities. This is expected to significantly reduce the hardware cost of the perception module, providing key technical support for automakers to incorporate advanced intelligent driving functions while controlling overall vehicle prices, thus accelerating the industrial progress of intelligent driving.

[0049] This invention is expected to promote the widespread adoption of active safety features, automatic emergency braking, lane centering, and high-speed NOA in a wider range of vehicle models, improving overall road traffic safety and efficiency. Simultaneously, its lightweight and low-power characteristics help reduce the energy consumption and heat dissipation burden of onboard computing units, indirectly reducing overall vehicle carbon emissions. Most importantly, it does not rely on expensive foreign sensors, LiDAR, or closed algorithm platforms; the core modules can be implemented entirely based on domestically produced chips and the open-source framework MMDetection3D. This will help build an independent and controllable intelligent vehicle supply chain system, enhancing my country's technological influence and industrial competitiveness in the global automotive industry transformation.

[0050] The following is for reference. Figure 6 It shows a schematic diagram of the structure of a computer system 600 suitable for implementing an electronic device according to embodiments of the present invention. Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0051] like Figure 6 As shown, the computer system 600 includes a central processing unit (CPU) 601, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 602 or programs loaded from storage section 608 into random access memory (RAM) 603. The RAM 603 also stores various programs and data required for the operation of the system 600. The CPU 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0052] The following components are connected to I / O interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a liquid crystal display (LCD) and speakers, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card and a modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to I / O interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 610 as needed so that computer programs read from it can be installed into storage section 608 as needed.

[0053] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 609, and / or installed from removable medium 611. When the computer program is executed by central processing unit (CPU) 601, it performs the functions defined in the methods of the present invention. It should be noted that the computer-readable storage medium of the present invention can be a computer-readable signal medium or a computer-readable storage medium or any combination thereof. The computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this invention, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable storage medium other than a computer-readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0054] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Python, and C++—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0055] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0056] The modules described in the embodiments of the present invention can be implemented in software or in hardware.

[0057] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable storage medium carries one or more programs that, when executed by the electronic device, cause the electronic device to: acquire an image dataset, preprocess the dataset and train a high-performance teacher network; construct a student network using a four-fold collaborative compression strategy of backbone lightweighting, channel pruning, spatial resolution optimization, and interpolation operator replacement, and sequentially execute image feature extraction, depth estimation, BEV spatial projection, BEV feature encoding, and multi-task output steps, and perform knowledge distillation on the student network based on the teacher network; after the student network training is completed, evaluate and optimize the model performance through baseline performance testing and hyperparameter tuning, and verify the scene understanding performance of the student network using blind testing submission and result analysis, achieving lightweight and efficient perception while maintaining accuracy.

[0058] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention is not limited to the specific combination of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A lightweight method for monocular BEV perception for low-resource autonomous driving platforms, characterized in that, include: Obtain an image dataset, preprocess the dataset, and train a high-performance teacher network; A student network is constructed using a four-fold collaborative compression strategy of backbone lightweighting, channel pruning, spatial resolution optimization, and interpolation operator replacement. The student network is then sequentially executed with image feature extraction, depth estimation, BEV spatial projection, BEV feature encoding, and multi-task output steps. Knowledge distillation is then performed on the student network based on the teacher network. After the student network is trained, the model performance is evaluated and optimized through baseline performance testing and hyperparameter tuning. Blind test submission and result analysis are used to verify the scene understanding performance of the student network, achieving lightweight and efficient perception while maintaining the same accuracy.

2. The lightweight method for monocular BEV perception according to claim 1, characterized in that, The image feature extraction specifically includes: MobileNetV3-Large is used as the backbone network, and lightweight feature extraction is achieved through an inverted residual bottleneck structure, SE module, and Hard-Swish activation. A lightweight feature pyramid network containing only three layers (P3-P5) is then constructed thereafter.

1. Convolution for channel alignment, 3.

3. Depthwise separable convolution is used for feature fusion, and multi-scale features are uniformly upsampled to 1 / 8 resolution of the original image.

3. The lightweight method for monocular BEV perception according to claim 1, characterized in that, The depth estimation specifically includes: The number of depth prediction channels was reduced from 118 to 32 uniform depth intervals; the standard 3 in the depth branch was also reduced. Replace 3 convolutions with depthwise convolutions plus 1 A combination structure of 1-point convolution is used; the depth probability is normalized by Softmax along the depth dimension; the pixel features are copied into the 3D space voxel by weighting according to the depth probability based on the camera intrinsic and extrinsic parameters, realizing the geometric dimensionality upgrade from 2D to 3D.

4. The lightweight method for monocular BEV perception according to claim 1, characterized in that, The BEV space projection specifically includes: Three-dimensional voxels are average-pooled along the height dimension to generate two-dimensional BEV feature maps, with the resolution set to 100. 100; Nearest neighbor interpolation is used instead of bilinear interpolation in voxel-to-raster mapping; After projection, Layer Normalization is performed on the BEV features to normalize the channels.

5. The lightweight method for monocular BEV perception according to claim 1, characterized in that, The BEV feature encoding step includes: Use 1-2 layers of 3 A lightweight convolutional structure consisting of 3 convolutions, BatchNorm, and ReLU is used to encode BEV features. The number of channels is fixed at 128, and no Transformer, Deformable Convolution, or positional encoding mechanisms are introduced.

6. The lightweight method for monocular BEV perception according to claim 1, characterized in that, The multi-task output step specifically includes: Based on the same set of encoded BEV features, respectively through 1 1. Convolutional methods are used to construct a 3D object detection head and a semantic map segmentation head. The semantic map segmentation head is a lightweight FCN structure that outputs probability maps of semantic categories such as lane lines, sidewalks, drivable areas, and parking lines.

7. The lightweight method for monocular BEV perception according to claim 1, characterized in that, The knowledge distillation specifically also includes: constructing a two-path distillation framework for teacher network models and student network models, including: Feature distillation: Aligning intermediate representations of teachers and students in the BEV feature space; Output distillation: Align the predicted distributions of teachers and students at the task output layer to take advantage of the class relationships and uncertainty information implied by the soft labels in the teacher output.

8. The lightweight method for monocular BEV perception according to claim 7, characterized in that, The feature distillation specifically involves extracting the BEV feature maps of both teachers and students after the BEV encoder output. The feature map size is H. W C, calculate the L2 distance loss pixel by pixel, the specific formula is: in, The feature distillation loss is represented by H, where H represents the height of the BEV feature map, W represents the width of the BEV feature map, C represents the number of channels in the BEV feature map, and i represents the index in the height direction of the feature map. i H,j represents the index along the width direction of the feature map, 1 j W, k represents the index of the feature map channel direction, 1 k C, This represents the feature value of the student network model at position (i,j,k). This represents the feature value of the teacher network model at position (i,j,k).

9. The lightweight method for monocular BEV perception according to claim 7, characterized in that, The output distillation targets the heatmap and BEV segmentation probability map in 3D detection, using temperature-scaled KL divergence loss, with the specific formula as follows: Where p( =softmax(Y / ), representing the soft probability distribution after temperature scaling, and Y representing the original values ​​of the model. This represents a temperature parameter, a scalar value greater than 0, and 'c' represents the category index. This represents the probability output of the teacher network model in class c. This represents the probability output of the student network model in class c. This represents the output distillation loss, and KL divergence is used to measure the difference between the output probability distributions of teachers and students.

10. A computer program product, characterized in that, It stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1-9.

Images