Autonomous vehicles rely heavily on accurate and efficient scene segmentation for safe navigation and efficient operations.Traditional Bird’s Eye View(BEV)methods on semantic scene segmentation,which leverage multimo...Autonomous vehicles rely heavily on accurate and efficient scene segmentation for safe navigation and efficient operations.Traditional Bird’s Eye View(BEV)methods on semantic scene segmentation,which leverage multimodal sensor fusion,often struggle with noisy data and demand high-performance GPUs,leading to sensor misalignment and performance degradation.This paper introduces an Enhanced Channel Attention BEV(ECABEV),a novel approach designed to address the challenges under insufficient GPU memory conditions.ECABEV integrates camera and radar data through a de-noise enhanced channel attention mechanism,which utilizes global average and max pooling to effectively filter out noise while preserving discriminative features.Furthermore,an improved fusion approach is proposed to efficiently merge categorical data across modalities.To reduce computational overhead,a bilinear interpolation layer normalizationmethod is devised to ensure spatial feature fidelity.Moreover,a scalable crossentropy loss function is further designed to handle the imbalanced classes with less computational efficiency sacrifice.Extensive experiments on the nuScenes dataset demonstrate that ECABEV achieves state-of-the-art performance with an IoU of 39.961,using a lightweight ViT-B/14 backbone and lower resolution(224×224).Our approach highlights its cost-effectiveness and practical applicability,even on low-end devices.The code is publicly available at:https://github.com/YYF-CQU/ECABEV.git.展开更多
Bird's-eye-view(BEV)perception is a core technology for autonomous driving systems.However,existing solutions face the dilemma of high costs associated with multimodal methods and limited performance of vision-onl...Bird's-eye-view(BEV)perception is a core technology for autonomous driving systems.However,existing solutions face the dilemma of high costs associated with multimodal methods and limited performance of vision-only approaches.To address this issue,this paper proposes a framework named“a lightweight pure visual BEV perception method based on dual distillation of spatial-temporal knowledge”.This framework innovatively designs a lightweight vision-only student model based on Res Net,which leverages a dual distillation mechanism to learn from a powerful teacher model that integrates temporal information from both image and light detection and ranging(LiDAR)modalities.Specifically,we distill efficient multi-modal feature extraction and spatial fusion capabilities from the BEVFusion model,and distill advanced temporal information fusion and spatiotemporal attention mechanisms from the BEVFormer model.This dual distillation strategy enables the student model to achieve perception performance close to that of multi-modal models without relying on Li DAR.Experimental results on the nu Scenes dataset demonstrate that the proposed model significantly outperforms classical vision-only algorithms,achieves comparable performance to current state-of-the-art vision-only methods on the nu Scenes detection leaderboard in terms of both mean average precision(mAP)and the nu Scenes detection score(NDS)metrics,and exhibits notable advantages in inference computational efficiency.Although the proposed dual-teacher paradigm incurs higher offline training costs compared to single-model approaches,it yields a streamlined and highly efficient student model suitable for resource-constrained real-time deployment.This provides an effective pathway toward low-cost,high-performance autonomous driving perception systems.展开更多
基金funded by the National Natural Science Foundation of China,grant number 62262045the Fundamental Research Funds for the Central Universities,grant number 2023CDJYGRH-YB11the Open Funding of SUGON Industrial Control and Security Center,grant number CUIT-SICSC-2025-03.
文摘Autonomous vehicles rely heavily on accurate and efficient scene segmentation for safe navigation and efficient operations.Traditional Bird’s Eye View(BEV)methods on semantic scene segmentation,which leverage multimodal sensor fusion,often struggle with noisy data and demand high-performance GPUs,leading to sensor misalignment and performance degradation.This paper introduces an Enhanced Channel Attention BEV(ECABEV),a novel approach designed to address the challenges under insufficient GPU memory conditions.ECABEV integrates camera and radar data through a de-noise enhanced channel attention mechanism,which utilizes global average and max pooling to effectively filter out noise while preserving discriminative features.Furthermore,an improved fusion approach is proposed to efficiently merge categorical data across modalities.To reduce computational overhead,a bilinear interpolation layer normalizationmethod is devised to ensure spatial feature fidelity.Moreover,a scalable crossentropy loss function is further designed to handle the imbalanced classes with less computational efficiency sacrifice.Extensive experiments on the nuScenes dataset demonstrate that ECABEV achieves state-of-the-art performance with an IoU of 39.961,using a lightweight ViT-B/14 backbone and lower resolution(224×224).Our approach highlights its cost-effectiveness and practical applicability,even on low-end devices.The code is publicly available at:https://github.com/YYF-CQU/ECABEV.git.
基金supported by the National Natural Science Foundation of China(42476084,62203456,42276199)the Stable Support Project of National Key Laboratory(WDZC 20245250302)the National Key R&D Program of China(2024YFC2813502,2024YFC2813302)。
文摘Bird's-eye-view(BEV)perception is a core technology for autonomous driving systems.However,existing solutions face the dilemma of high costs associated with multimodal methods and limited performance of vision-only approaches.To address this issue,this paper proposes a framework named“a lightweight pure visual BEV perception method based on dual distillation of spatial-temporal knowledge”.This framework innovatively designs a lightweight vision-only student model based on Res Net,which leverages a dual distillation mechanism to learn from a powerful teacher model that integrates temporal information from both image and light detection and ranging(LiDAR)modalities.Specifically,we distill efficient multi-modal feature extraction and spatial fusion capabilities from the BEVFusion model,and distill advanced temporal information fusion and spatiotemporal attention mechanisms from the BEVFormer model.This dual distillation strategy enables the student model to achieve perception performance close to that of multi-modal models without relying on Li DAR.Experimental results on the nu Scenes dataset demonstrate that the proposed model significantly outperforms classical vision-only algorithms,achieves comparable performance to current state-of-the-art vision-only methods on the nu Scenes detection leaderboard in terms of both mean average precision(mAP)and the nu Scenes detection score(NDS)metrics,and exhibits notable advantages in inference computational efficiency.Although the proposed dual-teacher paradigm incurs higher offline training costs compared to single-model approaches,it yields a streamlined and highly efficient student model suitable for resource-constrained real-time deployment.This provides an effective pathway toward low-cost,high-performance autonomous driving perception systems.
