Target detection network based on multi level information fusion of radar and vision
In response to the problem of poor detection performance of some high-risk moving targets in autonomous driving perception tasks due to complex road environments and insufficient fusion of onboard radar and camera data,this paper designs an object detection network MLFusionNet that integrates radar and visual multi-level information based on Centerfusion.Firstly,data level fusion is added to the input layer,which concatenates the radar echo features with the image in the form of pixel values,and then inputs them into the encoding and decoding network through a secondary residual fusion module,enriching the input information of the network;then,a bottleneck structured context module was designed between the encoder and decoder of the backbone network,which obtains broader contextual information from the feature map through a multi branch convolutional structure and reduces the number of parameters through compression channels;finally,a parallel attention fusion module was designed to solve the problem of insufficient feature level modal fusion.The experimental results on the nuScenes dataset showed that the NDS of MLFusionNet reached 46.6%,which increased the mAP of cars,trucks,and pedestrians by 1.4、3.0 and 1.5 percentage points respectively compared to the multimodal network Centerfusion.This indicates that the network pays more attention to high-risk dynamic targets in the driving environment.
deep learningtarget detectionfeature fusionattention mechanism
万方数据
维普
