A DEEP HYBRID MODEL FOR CROWD COUNTING WITH UNCERTAINTY ESTIMATION VIA NORMALIZING FLOWS

Abstract

Even though counting people in crowded places is a demanding task in computer vision, it’s still highly valuable. Geography helps improve public safety, guide city planning and handle big events. The main issues are that in real life, traditional methods often fail since people’’ sizes in the frame (scale) vary, they get hidden (occlusion) and they might not be distributed evenly (non-uniformity). To deal with these issues, this thesis suggests a new deep learning system that combines two things: a reliable way to calculate crowd density and a method to measure the model’s confidence in its predictions. A ResNet-101 network is used as the foundation and a FPN is added on top to help the system detect and interpret people in groups from different positions and scales. Because of this arrangement, the model is better able to make density maps when scenes contain a lot of warping or uneven crowding. Here, the main novelty is including a Real NVP network that measures how reliable the estimates are. Essentially, global features are extracted by the network, then passed through a fully connected layer before going through RealNVP which converts them into a probabilistic state. The model estimates the trustworthiness of its predictions by judging the likelihood of such features under a normal distribution. To give this uncertainty true value, an additional loss function is introduced. With it, the model is only uncertain once it is likely to get the answer wrong. The method described in this paper is practical and offers interpretable results for crowd counting. It not only counts out the people in a shot, but also explains how certain it is about what it found. Transparent AI is an important move for making systems easier to explain. This approach can eventually be applied to video data, where tracking individuals over different times would be more valuable or to segmentation and object detection, both of which require dealing with uncertainty.