Pedestrian detection from surveillance camera

As the global cities and technology are growing quickly, the field about intelligent monitoring has gradually become one of the main topics. As a major issue in the field of intelligent monitoring, the problems of crowd counting gradually enter people's field of vision. In crowded scenes, counting issues are very important for safety and flow restrictions. In recent years, convolutional neural networks (CNN) have achieved outstanding results in the field of computer vision research. Its outstanding performance in image feature extraction and model generalization effectively solves the feature extraction problem of crowd counting under complex background. In view of the complexity of some scenes, the current neural network models for crowd counting use deeper and more complex structures to get the desired performance. In order to get more efficient methods of extracting feature maps, in this report we first analysis several typical models and their performance. Based on their drawbacks then propose a regression-based neural network including residual net and two types of attention module. The attention module is applied to both channel and spatial dimensions, which can improve the feature extraction of the network model without significantly increasing the amount of calculation and parameters. Besides, through using various-scale architecture, we can also get high-resolution density maps. Then simulation results on the dataset named ShanghaiTech and UCF_CC_50 show pretty good performance compared to a few previous works. At last, several actual problems and future research topics are presented in order to make this model more practical.Master of Science (Communications Engineering