An abnormal behavior detection method based on a linear dynamic system and a deep network

Abstract

The invention provides an abnormal behavior detection method based on a linear dynamic system and a deep network. An LDS dynamic system model and a TSN deep learning network are combined, the space-time feature information of actions is extracted through the TSN, then action features are connected in series through an LDS to form complete behavior features, and finally the behavior type is judgedthrough a support vector machine (SVM). According to the method, the 3D convolutional network is established on the basis of the residual network, and the 3D convolutional kernel is established in a 2D+1D form, so that the network parameter quantity is reduced, and the problem that the original 3D network cannot preload the weight is solved. A residual 3D network is introduced into the TSN structure, so that the feature extraction capability of the network is improved. The number of network layers is increased, and the fitting capability of the network is improved. According to the invention,the high-precision identification of long-sequence abnormal actions can be realized, and finally the accurate monitoring of abnormal behaviors is realized.

Description

technical field [0001] The invention relates to a video detection method, in particular to an abnormal behavior detection method based on a linear dynamic system and a deep network. Background technique [0002] In recent years, video surveillance has been widely used in park security, traffic monitoring, indoor monitoring and other public environments. With the popularization of monitoring, a fast and stable detection method is needed to analyze and process abnormal behaviors in videos. The so-called abnormal behavior refers to actions that do not comply with regulations or routines, that is, contrary to conventional behaviors, such actions are dangerous or hidden dangers. By detecting the abnormal behavior of people or crowds, key early warning information can be provided, and the harm caused by emergencies can be reduced from the source. [0003] Abnormal behaviors have the following characteristics: (1) The duration of the action is short, the features are not clear, a...

AI Technical Summary

Problems solved by technology

However, 3D-CNN has the following defects: 1) The high-dimensional convolution kernel makes the number of weight parameters of 3D-CNN much larger than that of 2D-CNN with the same structure, the model volume is large, and the calculation speed is low

2) Also due to the high-dimensional convolution kernel, the model cannot obtain excellent initialization weight parameters through pre-training, and it is difficult to train and fit

3) The number of network layers is too shallow, the nonlinear structure is too small, and the ability to extract high-dimensional spatiotemporal features is limited

[0015] The TSN network can extract time and space features in parallel, that is, two types of features are obtained from RGB images and optical flow images respectively. The type and range of space-time features extracted by it are larger than 3D-CNN and LRCN networks, but it has the following defects: 1) Relying on optical flow images for action feature extraction, the accuracy of optical flow directly affects the classification accuracy

However, high-precision optical flow images (such as bidirectional sparse optical flow) require a lot of computing resources, so the processing speed of the network is slow

2) TSN ignores the connection between time and space features during feature fusion, and only performs a simple merge operation, lacking consideration of the temporal relationship between time and space features

Although multi-layer convolutional layers are used for feature fusion in the literature [8], the improvement accuracy is limited

3) Due to the poor ability of 2D-CNN to extract temporal features, TSN cannot acquire temporal and spatial features in optical flow images, which also reduces the quality of fusion features

That is, a behavior is composed of multiple actions, and a single action can also form multiple actions, and it is impossible to accurately determine whether the behavior is abnormal through a single action segment

Images