| Analyzing the presence of abnormal behavior in video has a wide range of applications in intelligent transportation,smart cities,and security for major domestic and international projects.With the increasing national emphasis on security in recent years,analyzing abnormal behavior in video has become a popular research direction in modern computer vision.In this paper,the abnormal trajectory detection task is viewed as a classification problem,and the research is carried out based on multi-target tracking with the moving trajectory of the target centroid.Firstly,this paper designs a target detector for small objects in traffic scenes.Then,based on this,a Kalman filter and the apparent features of the object are used for data association to complete the multi-target tracking task and obtain the moving trajectory of the object,after which a sequence-to-sequence model is used to encode the trajectory,then a one-dimensional convolution-based feature extractor is used for feature extraction of the encoding,and finally the trajectory is classified.The main work of this paper includes.(1)A lightweight target detector for small objects in traffic scenes is proposed.Firstly,we try to add different combinations of attention modules on the skeleton network of the target detector,compare the performance of the target detector under different module combinations,and get the best combination scheme.Then the coupled detection heads in the original network are replaced using decoupled detection heads to improve the performance and pre-convergence speed of the target detection model.Finally,a feature fusion network for small targets in traffic scenes is designed,which has a minimum downsampling ratio of 4 for the input feature map and can retain the information of small targets well.Compared with the original model and the current mainstream lightweight target detection model,the detection capability of the target detection model designed in this paper has been greatly improved,especially the small target detection capability.(2)A Re-ID feature extraction network with attention module is proposed,and data correlation is performed using object apparent features and D-IoU metrics.The detection of surveillance video using the target detection model can get the target detection result of each frame,and then the feature extraction of the detection frame and the prediction frame obtained using Kalman filter is performed using the Re-ID feature extraction network proposed in this paper,and the cost matrix is obtained using the cosine distance and the Marxian distance,after which the cost matrix is matched using the Hungarian algorithm,and finally for the unmatched detection frames and trajectories,and then matching using the D-IoU cost matrix.By this design,the identity switching problem can be significantly reduced and the performance of the tracking model can be improved.Finally,compared with the original model,the indexes such as MOTA have been significantly improved.(3)A trajectory classifier based on sequence-to-sequence model and one-dimensional convolution is proposed.In this paper,the abnormal trajectory detection is viewed as a binary classification problem,and firstly,the trajectory is encoded using a sequence-tosequence model,which uses LSTM and Transformer,which can learn the dynamic time characteristics of the trajectory and avoid the problem of unfocused abnormal trajectories due to long time monitoring.Further,a 1D convolution-based classifier is designed,using 1D convolutional kernels of different sizes to encode sequences for feature extraction operations,where different sizes of convolutional kernels have different perceptual fields and can extract temporal features of different perceptual fields,and then the features are stitched together and adjusted by linear layers and fed into Softmax for classification output.In comparison with other methods,our proposed method is more robust,can migrate for different traffic scenes,and the accuracy and other indexes are better than other methods. |
PDF Full Text Request