Object Detection And Recognition With Applications To Traffic Scenes

Object detection and recognition show great significance in computer vision.It is a basic as well as tough process for advanced applications.This thesis concentrates on salient object detection in the absence of a goal-driven situation,as well as vehicle attributes detection,vehicle identification and retrieval,and vehicle logo recognition in traffic scenes that are driven by goals.The main contributions of this thesis are as follows:1)For salient object detection.Most existing saliency detection algorithms concentrate on obtaining good results for images with single salient object,while they produce poor generalization power when tested on more complex scene.In this paper,we present a novel framework to detect saliency in object-level through fusing objectness estimation into the process of salient object detection.Different from most existing methods that evaluate saliency via aggregation of adjacent pixels or regions,our approach peels background regions step by step until all the independent foreground objects are left.During peeling,each region's saliency,objectness and background are evaluated at each step.Instead of extracting from saliency map,the proposed method can obtain salient objects directly,and different salient scores can be assigned to different salient objects.Experimental results show that the proposed method is effective and achieves state-of-the-art performance in several benchmark datasets,especially on PASCAL_S and SED2 that offer salient objects in more complicated scenes.2)For vehicle attributes detection and vehicle re-identification.We firstly collect a vehicle dataset VAC21 that contains 7129 images of five types of common vehicles.Then we make best effort to label 21 classes of structural attributes hierarchically by bounding boxes.As far as we know,this is the first dataset with so many detailed attributes labeled.Based on this dataset,we adopt the state-of-the-art one-stage detection method SSD as a baseline model for detecting attributes at first.Then we make a few important modifications tailored for this application to improve accuracy.i.e.adding proposals from low-level layer to improve the accuracy of small objects;importing the focal loss to improve the mAP.The results of attributes detection can be widely applied into a series of vision tasks that focus on vehicles.Furthermore,based on deep convolutional neural network,we propose a novel method which the ROIs are structured as a unique identifier for vehicle re-identification and retrieval.Then a series of experiments are conducted on database VehicleID.The experimental results show that our method outperforms state-of-the-art methods.3)For vehicle logo recognition,we firstly build a vehicle logo dataset named ‘VehicleLogo197' which contains 50318 logo images with 197 different models of cars.To our knowledge,the proposed dataset is largest in size and contains the most classes.Based on the proposed dataset,a compact residual convolution neural network is trained to classify the logos.To verify the effectiveness of the network and the usefulness of the proposed dataset,we compare our model against one recent state-of-the-art VLR network as well as four standard networks,namely,VLRNet,AlexNet,SqueezeNet,VGG16 and ResNet50.We carefully evaluate different models by varying the input image size from 56 × 56 to 227 × 227.Experimental results show that our model outperforms the compared nets on all settings with different input image sizes.Besides,we also conduct vehicle logo verification,which combines deep features extracted from the classification model and Joint Bayesian to train a verification model.The experimental results the features extracted from classification model are also effective for verification.