The Design And Implementation Of An Object Detection System Based On Deep Learning For Images From Traffic Monitors

Object detection of images is to locate and classify objects in these images.Technology of object detection for images from traffic monitors to get traffic information is very import for the development and construction of safe city and intelligent traffic.Objects are splitted from images by obect detection,which can promote traffic behavior analysis and traffic crime tracking。At present,mainstream object detection algorithms,which can detect common objects quickly and accurately,are based on deep learning.But for images from traffic monitors,speed and accuracy can not meet the needs of traffic at the same time.Based on deep learning,a fast and accurate object detection system for images from traffic monitors is created.Objects in images from traffic monitors are mainly divided into 3 classes:car,cyclist and pedestrian.Traditional object detectors first generate proposals.After that features are extracted.Then a classifier on these proposals is executed.But the speed is slow and the accuracy is not satisfying.According to the cell idea on images from YOLO,and inspired by R-FCN,All the fully-connected layers of YOLO’s network are replaced with an average pool layer for the purpose of reproducing a new network.The loss function is optimized after the proportion of bounding coordinates error is increased.A new object detection method,named OB-FCN,which means object detection by full convolution network,is produced.It is 1.18 times faster than YOLO,while outperforming other region-based approaches like R-CNN in accuracy.To improve accuracy further,the combination of OB-FCN and R-FCN is applied.Two algorithms run separately and parallel.If the intersection of two bounding boxes in their outputs is greater than 0.4 times of total area of the two boxes,we take their intersection as output.We take the average class probabilities of the two algorithms’ outputs as class probabilities.The mAP(mean Average Precition)of OB-FCN+R-FCN can reach 83.4%on our test set while YOLO is 64%.For challenging images in nights,pre-processing is presented using the histogram equalization approach.A scheme based on pipeline is applied to make the process of night images faster.There are two chennals in our system.One is for day images and the other is for night images.And there are two algorithms in each channel.One is responding to high speed requirement and the other meets the need of high accuracy.