In automatic video surveillance applications,accurately identifying abnormal human behavior is a very difficult task.To solve the problem of efficient recognition of abnormal human activities in the monitoring system,an abnormal behavior recognition model ICBAM-ResNet50 that strengthens the fusion of local and global feature information is proposed.Experiments are carried out on the UTI and CASIA datasets,and the results show that the accuracy of the study is 7%and 8%higher than that of the ResNet50 model,respectively.The ICBAM module introduces one-dimensional convolution to replace the MLP operation of channel attention in the original CBAM,integrating local temporal features into channel descriptors.Which alleviates the problem of ignoring information interaction caused by global processing in the channel dimension.Secondly,the spatiotemporal attention mechanism is introduced to replace the single spatial attention mechanism in CBAM to improve the spatiotemporal representation ability of the model.Finally,the optimized CBAM module is embedded in ResNet50,and by pre-training it on ImageNet,the model achieves 98.8%and 97.9%accuracy on two benchmark datasets,respectively.Using the same dataset,the experimental results are compared with the original recognition method,and the results show that the model is superior to the other methods compared.
NETL
NSTL
万方数据
