Research On Target Anomalous Behaviour Detection Method Based On Feature Fusion

Anomalous behaviour detection technology has become an important scientific and technological tool to enhance the security of surveillance sites,which in real life can be prone to problems such as unclear judgement of the type of anomalous behaviour and poor detection accuracy.In order to remedy the above deficiencies,this paper investigates the anomaly detection models using supervised and unsupervised learning for different scenarios of anomalous behaviour,respectively,in order to improve the detection performance of the algorithm.The main content is as follows.(1)To address the problems of inaccurate definition of behaviour types and lack of manual markers in videos of abnormal behaviour collected in life,an attention and feature fusion algorithm for future frame prediction abnormal behaviour detection is proposed,which uses unsupervised learning to detect abnormal behaviour in video frames.Firstly,the downsampling of the underlying network has been improved,based on the U-Net network structure,to reduce the loss of information during pooling.Secondly,ECA model is used to activate classification features with different weights to enhance feature differentiation,reduce interference from irrelevant information and improve network performance.Then,in order to fuse the detailed features of the image,the BTE Net module is proposed for use in the U-Net deep encoder to enhance the dependency of the global information of the image.Finally,a new extracted feature module is constructed and added to the output part of the network feature map to better fuse the feature information of the context in the image and extract the key information that better fits the requirements as the output image.The discriminative network uses the Patch GAN structure,which allows the network to focus more on the detailed features of the image.The data show that the improved network has improved accuracy for anomalous behaviour,proving that the improved method has certain timeliness and robustness.(2)To address the problem of low accuracy in detecting anomalous behaviour at different scales with clear classification in real life,a YOLOv5 adaptive multi-feature neck fusion based anomalous behaviour detection algorithm is proposed.Firstly,for different scenes and different scales of anomalous behaviour,two self-constructed datasets are used to test the performance of the improved network.Secondly,the deep C3 module of the backbone network is replaced by a CSPT module incorporating a selfattentive mechanism to focus on the holistic nature of image features and strengthen the extraction of global information from the backbone network.Then,the neck network adopts an adaptive bi-directional feature pyramid structure,assigning different weights to the feature layers and adaptively fusing global semantic and texture information from the deep and shallow layers of the network to improve the flexibility of the network.Finally,a very small detection layer SH module is added to extract shallow small sensory field feature maps,and for small targets,network accuracy has improved.Meanwhile,solving the problem of difficult recognition of small-scale abnormal behavior(e.g.smoking).The data results show that for targets at different scales.The improved network not only improves the detection accuracy,but also the detection speed is relatively good,meeting the needs of detection in reality.