基于改进Transformer-无迹卡尔曼滤波器的智能车辆多模态3D目标检测方法

打开文本图片集

主题词：无人驾驶 Transformer 多传感器信息融合 多模态感知无迹卡尔曼滤波器 中图分类号：TK411 文献标志码：A DOI：10.19620/j.cnki.1000-3703.20241037

Design and Experimental Validation of an Enhanced Transformer-UK Multi-Modal 3D Object Detection System for Intelligent Vehicles

Zhang Zhening¹2，Liu Zhen12，Wang Huaqiang12 （1.PureElectric VehiclePower System DesignandTesting HubeiProvincialKeyLaboratory，Hubei UniversityofArtsand Sciences，Xiangyang 441o53; 2.Schol ofAutomotiveand Transportation Engineering，Hubei Universityof Arts and Sciences，Xiangyang 441053）

【Abstract]To enhance the accuracyof environmental perception andobjectdetection for intelligent vehicles，this paper proposes anovelfeaturefusion-based3Dobject detection method.First，the Transformer modelis improvedbyleveraginga multi-headself-attention mechanism tofullycapturelocal spatial feature informationwhile preservingdiversefeature weights， enablingtheextractionof pointcloudandimage features.An Unscented KalmanFilter（UKF）isthenutilized todesigna multimodalfusionsystem thatcombines image and pointclouddata，ultimatelyachieving 3Dobject detection.The proposed modelistrainedandtestedonboth publicandprivatedatasetsandiscomparedagainstseveral state-of-the-artalgorithms. Experimentalresultsdemonstratethattheproposedfeature fusion-based3Dobjectdetectionmodelachievesanimprovement in meanAverage Precision （mAP）of O.34 percentage points and3.03 percentage points compared to widelyusedalgorithms suchasF-ConvNetandF-PointNet，respectively.Forvehicleandcyclistdetection tasks，themAPincreasesby 2.52 percentage pointsand9.32 percentagepoints，respectively.Moreover，theinferenceresultsonreal-world vehicledataalign closelywith thetraining evaluationoutcomes.These findings indicatethattheproposedalgorithmimproves theaccuracyof objectdetectioninperceptionsystems toacertain extentand providesavaluablereference forthedesignof objectdetection systems in autonomous driving.

KeyWords：Autonomous driving，Transformer，Multi-sensor informationfusion，Multimodal perception， Unscented Kalman Filter （UKF）

【引用格式】张哲宁，刘祯，王化强.基于改进Transformer-无迹卡尔曼滤波器的智能车辆多模态3D目标检测方法[J].汽车技 术，2026（2）： 39-47. ZHANG ZN，LIUZ，WANGHQ.DesignandExperimental Validationof anEnhanced Transformer-UKFMulti-Modal3D ObjectDetectionSystemfor IntelligentVehicles[J].Automobile Technology，2O26（2）：39-47.

1前言

近年来，无人驾驶汽车环境感知技术发展迅速，目标检测作为环境感知的基础性关键环节备受关注。（剩余12194字）