Research On Object Detection And Semantic Segmentation In Driver Assistance System

With the development of society and the continuous advancement of artificial intelligence technology,the development and production of intelligent vehicles equipped with driver assistance system have become the future development direction of the automotive industry.Driver assistance system can effectively improve driving safety and protect people’s lives and property.Object detention and semantic segmentation are very important sensing methods in driver assistance system,so it is of great significance to study them.Recently,deep learning has achieved great success in the field of computer vision.Therefore,we decide to use deep learning technology to implement related perception algorithms.In terms of object detection,this dissertation proposes a new object detection network based on multi-scale features.The model is based on RetinaNet and introduces three different modules(highly parallelized feature pyramid module,feature enhancement module,and head subnetwork based on multi-scale information fusion).These three modules enhance the performance of different parts in the model.As a result,the detection capability of the whole model is also improved.In terms of semantic segmentation,this dissertation implements a semantic segmentation network based on contextual information modeling.This model uses ResNet as the backbone network and uses the joint pyramid upsampling module to extract the feature map for subsequent segmentation.Finally,The feature map is input to the head subnetwork to obtain the final segmentation result.In this dissertation,two head subnetworks that can model context information are designed for semantic segmentation,which are respectively based on multi-scale contextual information and attention mechanism.Meanwhile,this dissertation also proposes a multi-task model based on the above two models.This model reduces redundant calculations at the lower layers of the network,which can speed up system processing and reduce memory usage.