The Study Of Object Detection Methods Via Deep Learning In Urban Transportation

Object detection is a critical point and challenging problem in ADAS(Advanced Driver Assistance System).Currently algorithms of object detection can be classified into two categories:IPBMs(image processing-based methods)and DLBMs(deep learning-based methods).IPBMs use some low-level features such as color?texture and edge contour to detect object regions or boundaries in the traffic scene image.However it doesn't work very well due to the complexity in real traffic scenario.Many factors,such as light,rain,and occlusion,could affect the performance.Compared to IPBMs,DLBMs are more reliable and robust.It can utilize the change of data appearance to increase the diversity of the samples.The more variety of object samples in urban traffic scene,the more accurate the object detection model will be.So as a branch of machine learning,deep learning has been widely used.But DLBMs still have many defects,such as long processing time when dealing with complex scenarios,low detection accuracy for small objects and so on.Our work aims to solve the detection accuracy issues(due to the convolution scale size)in RCNN(Regional Convolution Neural Networks).We propose a multitask convolution neural network based on the object localization algorithm.It has several contributions:1)different tasks share the same feature extraction layer in the feature extraction stage.2)the corresponding sub-network for different levels of information training works respectively.3)the accurate target location is obtained by a fusion of the high-level and low-level information.The experimental results in PASCAL VOC 2007 database and traffic scene database show that the algorithm can effectively improve the accuracy of object localization.At the same time this paper also proposes a rapid end-to-end convolution neural network,which is used in the lane line and traffic markers to detect and locate small targets.Besides,it adjusted the network perception domain while added the changes of spatial domain to the generation mechanism of candidate box on the basis of VGG16 network.With the small target candidate box(such as lane line block,etc.),the lane line equation can be obtained from the optimization method.Finally,we verified the performance of the algorithm on KITTI-ROAD data set and traffic scene data set.In terms of small target lane line blocks,the average accuracy has reached up to 64.3%using the method we mentioned in this paper,which is significantly higher than other algorithm performance.Finally,this paper proposes a convolution neural network based on hourglass cascade structure for traffic sign detection.This algorithm is on the basis of the residual network,which combined with SSD network framework,funnel structure with deconvolution operation,high-level semantic information and low level location information.Meanwhile residual block is utilized in detection module to optimize the training process.At last,we test the performance of the algorithm on two traffic sign data sets,TT100K and GTSDB.While the algorithm accuracy failed to reach the advanced level,but compared with other algorithms(DSSD,Faster RCNN),the performance has obviously improved.