Human detection from surveillance camera

Abstract

Surveillance cameras are widely installed along roadways. With the help of object detection algorithms, it has become easier to monitor a large number of areas, detect threat and other abnormalities. This in turn will increase the effectiveness in response towards abnormal situations. This project will focus on the Single Shot Multibox detector as a method to detect humans from images captured from surveillance cameras. The model is trained with a dataset made up from images from surveillance cameras placed in Nanyang Technological University area. This project hopes to develop a model which can be used to detect humans and count the number of people in the area. With the help of Keras port of the Single Shot Multibox detector in the implementation of the model, the training and evaluation has been made simpler. A total of 11 models will be trained and evaluated on their average precisions. In addition, 3 of the models will then be evaluated on their performance in real time setting.