Research On Multi-Source Data Fusion Target Detection Technologies For Intelligent Driving Scenarios

As artificial intelligence technology continues to develop,the urban transportation industry is gradually moving towards intelligence and automation,in which the intelligent driving system plays an important role.As an important component of intelligent driving systems,target detection algorithms provide strong support for the safe operation of vehicles.However,due to the dynamic nature of intelligent driving scenarios and the diversity of targets to be detected,achieving accurate detection and safe decision-making still faces serious challenges.Target detection technology based on single-source data cannot meet the demand for accurate target detection in complex and diverse real scenarios.Therefore,target detection technology based on multi-source data fusion has become the mainstream research direction.To achieve precise and effective target detection in intelligent driving scenarios,Two three-dimensional target detection algorithms based on multi-source data fusion are proposed in this thesis.The main work is as follows:Ⅰ.Firstly,focusing on the target detection technology,an overview is provided for target detection technologies based on different data sources;then,a detailed introduction is given for 3D object detection based on deep learning,point cloud-image spatiotemporal joint calibration,and point cloud-image fusion.Ⅱ.A dual-weighted mapping-based weather scene adaptive 3D object detection algorithm is proposed.Firstly,in response to the problem of insufficient fusion of existing multi-source data,a dual-weighted mapping image feature mapping algorithm is proposed,which maps the more detailed image features to the point cloud data based on the projection point position of the point cloud,obtaining point-by-point image mapping features containing image semantic information.Then,a weather scene-driven 3D object detection algorithm is proposed,which adaptively implements global feature fusion of point-by-point image mapping features and original point cloud features by calculating weather induction factors and scene scale factors,thereby obtaining more representative and robust fusion features,and inputting the fusion features into a 3D object detection network to obtain the final detection results.Experimental results show that the proposed algorithm can achieve accurate detection of targets in intelligent driving scenarios.Ⅲ.The multi-region fusion road scene adaptive 3D object detection algorithm is proposed.Firstly,in response to the high cost of existing multi-source data fusion,a multi-source feature fusion algorithm based on image regions is proposed.After generating the point cloud region features,the features of point cloud regions at different depths are weighted and fused based on an attention mechanism,and then the fused point cloud region features are combined with image region features to obtain the global region depth fusion features.Then,a road scene-driven 3D object detection algorithm is proposed,which adapts to the global region depth fusion features by road prior information(road category,congestion level)to achieve multi-depth resolution feature fusion and obtain scene-based fusion features.The fusion features are then input into the SSD network to complete 3D object detection.Experimental results show that the proposed algorithm balances speed and detection accuracy in intelligent driving scenarios and has certain application value.