Research On Object Detection For Vision Perception

Object detection is a popular research field in computer vision,and has been widely applied in the scenes of video surveillance and automatic driving.However,facing complex scenes,object scale variety and occlusion between objects,the current object detection methods based on Convolutional Neural Network(CNN)are not satisfied.To solve the above problems,this paper studies the object detection methods based on convolutional neural network.The main research results are as follows:(1)This paper improves Single Shot multibox Detector and presents a detection framework based on feature fusion(SingleNet).The method maps multi-layer convolutional neural network features into the same dimension and fuses them to enhance feature representation ability.In order to estimate the boundary box better,this method uses a set of dense anchors with multi-scale and multi-aspect ratio to regress the object boundary box.In addition,SingleNet can be easily deployed into object detection systems.The mean Average Precision(mAP)of SingleNet on PASCAL VOC2007 dataset is 0.776,which demonstrates the advantages of the proposed method compared with similar methods.(2)This paper proposes a Dual shot Face Detector(DSFD).This method introduces a feature enhancement module to generate the dual layers of the original feature layers to increase the semantic information contained in the image.For the original feature maps and the advanced feature maps,this method uses two set of anchors to compute progressive anchor loss,which can accelerate the model convergence.In order to better initialize the regressor,DSFD introduces a new data augmentation algorithm and anchor design strategy into anchor matching.On the WIDER FACE dataset,the mAP of DSFD is easy: 0.966,medium: 0.957,hard: 0.904;on the FDDB dataset,the mAP of DSFD is discontinuous: 0.991,continuous: 0.862.The results demonstrate the superiority of DSFD over current popular face detection methods based on convolutional neural network.(3)This paper presents an integrated framework for region proposal based object detection methods.Region proposal is adopted to re-sample feature maps to achieve better detection performance than SSD.This paper introduces three integration pattern of region proposal based object detection methods.The first method uses Non-local Maximum Suppression(NMS)to combine detection results of multi region proposal based object detection methods to improve the confidence of bounding box;the second method cascades the detection steps following feature maps sampling to adjust the quality of bounding box;and the third method is to ensemble last cascade detection steps following feature maps sampling,which can effectively improve the detection performance without increasing computation.In addition,facing the problems of limited training data and object scale variety,the framework augments training data by regular crop of training image.On the BDD100 K dataset,the mAP of region proposal based object detection integrated framework is 0.3411.In the 2016 Ucar Self-driving deep learning competition,the mAP of our framework was 0.829 and won the first place in the competition.