Research On Outdoor Scene Understanding Using Deep Convolutional Neural Networks

Scene understanding,as a challenging research hotspot,has been widely applied in many fields,such as,robot navigation,web search,security monitor and medical health.Each branch of scene understanding,including object detection and image semantic segmentation,has made breakthroughs,but there are still many deficiencies.For example,it is still difficult to capture invariable and robust features from moving targets for classification,on account of target deformation and external disturbance.Deep convolutional neural networks(DCNN)have achieved breakthroughs in scene classification by end-to-end feature leaning.But accurate image semantic segmentation using DCNN is still challenging.In response to above problems,the main research contents in this thesis are as follows:Firstly,an approach for moving object classification using multi-task spatial pyramid pooling deep convolutional neural networks is presented.Firstly,Gaussian mixture model is used for motion detection and accurate contour of moving target is obtained with image morphology post-processing.Then target image patch is extracted and input to multi-task spatial pyramid pooling deep convolutional neural networks for classification,and the semantic label is obtained.Experiment results showed that higher convolution features are robust to partial occlusion,overlap,and viewpoint change.High classification accuracy and sound semantic label about moving target can be obtained using multi-task spatial pyramid pooling deep convolutional neural networks.Secondly,to overcome the deficiency in robustness and representation of hand-crafted features,a new region-level scene labeling approach is proposed,which combines DCNN with the MeanShift segmentation algorithm.For each image,it is firstly segmented into local regions using the MeanShift algorithm.Then a deep convolutional neural network trained with images of target objects is employed to get the probability scores of randomly cropped samples of each segment.Object category of each local segment is finally determined by the average probability scores of its samples.We thoroughly analyzed the effects of some factors,including DCNN kernel size and number and data augmentation of training data,on the labeling results.Experiment results showed,comparing with SIFT based SEVI-VOVW model,this method is more superior in both accuracy and speed.At last,based on DCNN,a new scene understanding approach is presented that combines target detection and scene labeling.This method,when combined with background semantic segmentation method using HOG(Histogram of Oriented Gradients)texture feature and Support Vector Machine(SVM),is applied to robot navigation in campus environment.Faster R-CNN algorithm and DeepLab-CRFs model are used for foreground target's detection and semantic pre-segmentation,respectively.GrabCut algorithm is employed to get refined semantic segmentation about foreground target by integrating the results of Faster R-CNN algorithm and DeepLab-CRFs model.Experiments showed that this method can detect and segment target accurately and comprehensively,and can be effectively applied to robot navigation.