Research On Tunnel Object Detection Based On Convolutional Neural Network

Detecting tunnel objects by monitor video image can analyze traffic status in real time and reduce safety accidents,which is of great significance in ensuring tunnel traffic safety.The main purpose of the tunnel object detection algorithm is to locate and classify the object accurately.At present,most tunnel object detection algorithms locate objects by background subtraction,and classify the object by machine learning with artificial feature.The image classification algorithm based on machine learning depends heavily on the artificial features.It has a good effect on single category classification,but is difficult to deal with multiple classification tasks.Object detection algorithm based on Convolutional Neural Network(CNN)is developed rapidly and is good at dealing with multiple classification tasks.However,there are many difficulties in applying algorithm to tunnel scene.If it gets candidate regions at first and then classifies the regions by CNN,a large number of feature extraction operations will increase the time complexity of the algorithm,and can not satisfy the real-time requirements in tunnel.If it extracts image feature at first and then divides the candidate regions,the feature extraction of the whole image will lose a lot of low-level image information of the candidate regions and reduce the detection precision of the low resolution objects and small objects with long distance.Aiming at the problems above,this thesis proposes a tunnel object detection algorithm based on background subtraction and CNN.The basic thought of the algorithm is to extract candidate regions and get the positions of the objects by background subtraction,and then classify the candidate regions,determine the object class and remove redundant windows by a CNN model.The algorithm shows good performance under the condition of tunnel environment.The main research content of this thesis is divided into three parts.In the first part,the separating evaluation of color distortion is used to improve the foreground segmentation of the ViBe background modeling,and the C-ViBe algorithm is proposed.The result of the foreground segmentation of the C-ViBe algorithm has been optimized and the foreground object positions is recorded as candidate windows.The second part has discussed the performance of CNN at first.The relationship between the performance of neural network and the number of network layers is explained in theory.Then,the ReLU-S activation function has been proposed and the weight initialization strategy have been optimized based on the shallow CNN.Finally,a deep residual network ResNet34-mix with three fixed map features is proposed.Experiments show that ResNet34-mix has the characteristics of higher precision and faster speed compared with other networks,and it has practical application value.The third part proposes a tunnel object detection algorithm based on background subtraction and CNN.Experiments show that the tunnel object detection algorithm based on background subtraction and CNN has higher precision and faster speed compared with the traditional tunnel object detection algorithm and the object detection algorithm based on deep learning.