Research On Object Detection Methods Using Region-based Convolutional Neural Network

Object detection is an essential part of automated driving and driver assistance.It can assist intelligent vehicles in environmental perception and navigation system,and especially provide dependable outputs to avoid accidents or to guide vehicles drive safely.With the development of deep learning,accuracy of object detection has been greatly improved by using deep convolutional features.Compared with hand-crafted features,the deep convolutional features can be automatically learnt and extracted by deep convolutional neural network.However,the real-time performance of object detection algorithm based on deep convolutional network is restricted by the number of regions of interest,and a large amount of background regions in the regions of interest increase detection time-consuming.In addition,features of the topest convolutional layer are usually used to classify.But information loss due to multiple down-sampling and convolution of images will decrease the final detection precision to a certain extent.This paper focuses on the research of object detection process and convolution neural network framework.Algorithms of this paper will be used for pedestrian detection and vehicle detection.What the paper contributes to are listed below:1.In order to reduce the proportion of the background area in ROIs,a pre-classification algorithm for object detection based on Deepbox is proposed.Aiming at the three steps of object detection algorithm flow: ROIs extraction,feature extraction and target classification.The main idea of this algorithm is to exclude background regions from the extraction of ROIs and preserve the target area as much as possible.It can improve the recall rate of ROI and reduce the total number of ROIs.The number of ROIs directly affects the total time spent on object detection.The reduction of the total number of ROIs helps to improve the real-time performance of object detection algorithm.The algorithm is tested in the Pascal VOC 2007 database and compared with the mainstream algorithms.Experiments prove that the algorithm can effectively reduce the detection time and maintain a high detection accuracy.2.In order to decrease information loss of image convolution and down-sampling,an improvement has been made in the widely used convolutional neural network framework.The proposed algorithm called FusionCNN is based on multi-layer features fusion.This algorithm can solve information loss of multiple convolution and down-sampling and improve detection accuracy.This algorithm is effective to improve object detection accuracy of small and medium size objects in images.The experimental results show that the algorithm can effectively improve the accuracy of object detection in the Pascal VOC 2007 database,especially in pedestrian detection and vehicle detection.3.The Deepbox algorithm and FusionCNN algorithm are combined to carry out pedestrian detection and vehicle detection experiments in a database gathered by intelligent vehicles.The structure and parameters of the network are adjusted according to the characteristics of real vehicle images.In the experiment,compared with ACF algorithm and Fast-rcnn algorithm,it is proved that the proposed algorithm can reduce the detection time and improve the detection accuracy.