| Recently,the repaid growth of traffic information makes it impossible to review and process information manually.To meet the demands of traffic department,artificial intelligence is needed to monitor the traffic situation,raise warning when there is wrong with the road transport.The categories of anomalies are numerous and it is difficult to census them.In addition,it will take a lot of manpower and financial resources to find enough road traffic images or videos with anomalies in massive traffic information data.We can’t obtain enough training samples for each kind of anomalies.Under these circumstances,we have to face following challenges such as anomaly positioning,insufficient training data,overfitting problem or difficulty in convergence.In view of the above problems,this thesis studies and implements the anomaly recognition algorithm based on deep learning.Firstly,in order to generate the background model that is more suitable for road surveillance videos,we propose a background modeling method based on spatial information and temporal information.The pixels are clustered by k-means algorithm,and then the categories of pixels in the continuous video frames are classified as background or foreground by improved KNN method.After that,we find the positions of the anomalies.Experimental results show that the method can extract anomalies,ease the difficulty of locating anomalies.Then,the thesis proposes channel attention model for Siamese architecture to classify the anomalies that have been acquired.In the classification process,the category word vectors are embedded into the feature maps by attention layer.This method increases information contained in the feature map and makes it easier to measure the similarity.Aiming at the problem of lack of labeled data,we perform a two-stage training with transfer learning mode.By putting soft labels generated in the first stage into use,we enlarge the data volume and improve the performance of the model.The experimental results show that the classification accuracy of the proposed algorithm can reach above 95%,which is higher than the existing anomaly classification methods such as Auto Encoder.Finally,we propose a multi-perspective specific vehicle re-identification network to process and classify vehicle images.The network is composed of three subnetworks: vehicle perspective classification network,the multi-perspective feature extraction module(including adjacent perspective feature inference network and the vehicle key-point detection network)and multi-view global feature network.Apfi Net generates multiperspective hybrid features and generates multi-perspective global features via Mvgf Net and the positions of key points of vehicles are transformed into key point feature vector by multilayer perceptron in Mvgf Net.Vehicle reidentification is completed by calculating the Euclidean distances between features and sorting the distances.The experimental results show that the accuracy of Rank1 and Rank5 was 64% and 90%,both of them are higher than the existing reidentification methods such as FACT.The algorithm in this thesis can be applied to tasks such as detection of road anomalies,detection of abnormal behaviors of vehicle and vehicle tracking under multiple cameras,etc.It can play a role in traffic management,risk assessment,safety assurance and other aspects. |
