Multi-object Detection And Segmentation In Urban Traffic Scene Based On Mask R- CNN (FFM) Model

In recent years,urban traffic problems are deteriorating with the improvement of urbanization and mechanization level,which leads to much more demand of intelligent transportation and autonomous driving technology day by day.Environmental perception and urban traffic scene understanding are the key to autonomous driving technology,which can be actually summarized as multi-object detection and segmentation task in the wild.However,several drawbacks of current researches on multi-object detection and segmentation include data simplicity,insufficient feature representation,unsatisfactory model performance(especially for small object)and so on.The main research work in this dissertation is referred to as deep learning-based environmental perception and urban traffic scene understanding for autonomous driving,including multi-object recognition,detection and segmentation and so on.Firstly,based on the core idea of GANs model and its applications in image generation,W-DCGANs(Wasserstein-Deep Convolutional Generative Adversarial Networks)model is proposed in this dissertation,which can be used to generate highquality urban scene image to enrich data,as well as providing the basis for subsequent model train and evaluation experiments.Secondly,a novel model referred to as Mask R-CNN(FFM)is proposed to enhance convolutional feature extraction and simultaneous object detection and segmentation,which mainly introduces convolutional bone and FFM(Feature Fusion Module)architecture in the phase of feature extraction.This architecture helps to learn richer feature representations and largely alleviate information missing problem during feature mapping by hierarchical feature fusion and multi-scale feature pyramid transformation,which can boost better detection and segmentation performance.And then,case study on multi-object detection and segmentation in urban traffic scene is conducted to measure model performance and robustness ability,by using augmented autonomous driving dataset and proposed Mask RCNN(FFM)model.Extensive experimental results demonstrate Mask RCNN(FFM)model can achieve considerably competitive performance on multi-object detection and segmentation in urban scene,especially boosting model performance on small-object detection and segmentation when introducing convolutional bone and FFM architecture.Moreover,data augmentation not only helps model training and evaluation,but it also effectively improves model robustness performance.In conclusion,research in this dissertation is quite effective and practical: dataset augmentation by W-DCGANs model not only provides the basis of data for model train and evaluation experiment,but it also boosts the robustness and generality of model;environmental perception model Mask R-CNN(FFM)can achieve quite competitive multi-object detection and segmentation performance.This research is expected to be beneficial to the future development of autonomous driving technology and relevant research areas.