Crowd Counting Using Deep Learning

Abstract

Crowd counting is a challenging task, due mainly to the severe occlusions between dense crowds. This work aims to take a broader view to address crowd counting from the semantic modelling perspective. Crowd counting is essentially a function of pedestrian semantic analysis involving three key factors: pedestrians, heads, and their context structure. Different body parts information is an important clue to help us judge whether a person exists at a given position. Existing methods usually perform crowd counting from the perspective of directly modeling the entire body or the heads only, without explicitly capturing the composite body part information on semantic structure that is critical to crowd counting. In our approach, we first formulate the key factors of crowd counting as models of semantic scene. Then we convert the problem of crowd counting into a problem of multi-task learning, so that the models of the semantic scene are turned into different subtasks. Lastly, in a unified scheme the deep convolution neural networks (CNNs) are used to learn the subtasks. In terms of pedestrian semantic analysis, our approach encodes the semantic nature of crowd counting and provides a new solution. Our methodology outperforms state of the art approaches in experiments on four benchmark crowd counting datasets: 1. Sets of synthetic data 2. Databases-that's right Real-time data sets. The knowledge about semantic structure is proved to be an effective cue in crowd counting scene.