Improved Person Counting Performance Using Kalman Filter Based on Image Detection and Tracking

People detection and counting is becoming of great relevance in many application fields which range from the video surveillance of shopping centres to the management of public services such as bus stops. Various computer vision applications exploit Deep Learning technology for people recognition which recently showed great achievements. A very important task along with object/people detection is their tracking which is of fundamental importance especially when counting has to be performed. In this research work, a specific application for people counting at a bus stop is proposed, taking advantage of NVIDIA DeepStream SDK 5.0 detection and tracking algorithms. The final number of people is obtained through a Kalman Filter which aims at avoiding occlusion issues which are typical of people detection-tracking applications. The experimental results show the effectiveness of this approach in three scenarios with different complexity, significantly reducing the counting error with respect to pure detection counting.