A Deep Neural Network Approach for Top View People Detection and Counting

Abstract

People detection and counting is considered as one of the important application in video surveillance. Various computer vision and deep learning based methods have been developed which aim to provide efficient and accurate people detection/counting results using frontal view data sets. Furthermore, there are many challenges which occurs while detecting people including occlusion, perspective distortion, variations in human body pose, size and orientation, these challenges effect the results of developed detection and counting models. In this work, a deep neural network approach i.e. SSD (Single Shot multi-box Detector) is explored for people detection and counting. SSD model is used for detection and counting of people from significantly different viewpoint i.e. top view. To the extent of our knowledge, this is the first attempt to use deep neural network based model for top view people detection and counting. Furthermore, the impact of frontal view trained SSD model on top view test images is also discussed. The experimental results show the effectiveness of deep learning model by achieving promising results with average TPR of 95% and TPR 94.42% for indoor and outdoor environments respectively.