Person detection from surveillance camera

Abstract

Surveillance is a part of security. In most cases, this work costs a lot of time for people to observe and discover abnormal behaviors. Fortunately, the camera is the tool that can assist us to do this time-spending work. Compared to humans, the machine has a better vision. However, there is a doubt whether the detection performance of the machine is as accurate as that of human-being.

The accuracy of image classifiers has exceeded the human’s vision system. Due to the rise of deep learning, object detection from the surveillance camera can be realized. The video can be processed frame-by-frame as an image.

In this dissertation, the deep learning and the convolutional neural network (CNN) is introduced as the fundamental knowledge for image processing. And then, an algorithm for pedestrian detection has been proposed based on the Center and Scale prediction. A new dataset called WiderPerson, which has a wide range of diversity and density, is selected for training. Finally, experiments have been conducted to realize the Center and Scale predictor to detect pedestrians from images, even dealing with above hundreds of pedestrians in one image. The experiments also validate that this algorithm is effective for small-size objects.