Counting crowds using a scale-distribution-aware network and adaptive human-shaped kernel

Intelligent bus system plays a key role in the modern smart city. The number of passengers in the buses or at the stations is necessary for making an optimal scheduling policy of public buses. We develop a crowd counting algorithm to provide the counting information for a bus dispatch system in a human-machine system. In consideration of the challenges (e.g., pedestrian occlusions, non-uniform crowd distributions, and scale variations) existed in hand-crafted features based crowd counting, a scale-distributionaware multi-column convolutional neural network (SDA-MCNN) is presented to count crowds by summing up the output (denoted as the density map) of the SDA-MCNN. The SDA-MCNN is robust to scale variations by processing a crowd image with multiple convolutional neural network (CNN) columns and minimizing the per-scale loss. A weighted Euclidean loss is proposed to handle non-uniform crowd distributions. The loss can increase activations in dense regions and restrain activations in backgrounds. A new approach to estimate perspective maps of dense crowds is put forward to offer necessary information for generating density maps with human-shaped kernels. Evaluations on benchmarks are performed with other state-of-the-art counting approaches using deep neural networks. Comparative results verify the accuracy of our counting approach in challenging crowds. Evaluations on the real world BUS data reveal the accuracy of the proposed approach in counting passengers in spite of the complex environment. (C) 2019 Elsevier B.V. All rights reserved.