Researchers from Shanghai University of Electric Power developed a combination of atrous and fractional stride convolutional neural network
Automatic crowd analysis is increasing becoming a major part of current intelligent surveillance systems. Such approach can help to prevent severe accidents by offering information about the number of people and crowd density in a scene. Crowd counting and analysis can be used in video surveillance, traffic monitoring, public safety, and urban planning. The current intelligent surveillance system has several drawbacks such as inability to process a large-scale crowded environment with severe occlusion and non-uniformity. Now, a team of researchers from Shanghai University of Electric Power’s School of Electronics and Information Engineering developed a smart camera-aware crowd counting system.
The new system can operate through dilated convolutions and fractional stride convolutions. Atrous convolutions are focused on enlarging receptive fields. Such benefits can be used to add abundant characteristics into the network to enable the model to learn discriminative features that are relevant worldwide. This in turn can offer large count variations in the dataset of the system. The team used fractional stride convolutional layers as the back-end to restore the loss of details due to max-pooling layers in the earlier stages. This in turn allowed the team to revert on full resolution density maps. According to the researchers, the model structure has moderate complexity and robust generalization ability and the density estimation performance in densely crowded scenes was satisfactory.
The results were concluded from experiments on multiple datasets. The team adopted multi-task learning to enhance the loss function of crowd counting. Experimental comparison were conducted to demonstrate that the system offers better results. The team also verified the feasibility and effectiveness of the method. However, the system also has some drawbacks that are needed to overcome. An ideal alerting system focuses both on accurate crowd counting and predicts crowd behavior. Moreover, the system also needs to evaluate local information such as sub-group trajectory analysis of the crowd. In further research, the team plans to focus on compressing the broad network to include different real-time embedding systems. The research was published in the journal MDPI Sensors on March 18, 2019.
Subscribe to our newsletter to get notification about new updates,information, discount, etc..