Progressive Crowd Enhancement De-Background Network for crowd counting

Crowd counting is a very difficult task due to the presence of cluttered backgrounds in crowd scenes. Although recent counting algorithms have achieved great progress, most of them may inadvertently cause the loss of crowd information when removing background noise from images. To remedy this problem, we propose a simple but effective architecture, called Progressive Crowd Enhancement De-Background Network (PCEDNet), to minimize background interference and produce high-quality density maps. Specifically, we construct a Crowd Enhancement De-Background Module (CEDM) that employs the Parametric Rectified Linear Unit (PReLU) to retain more information from negative feature values and utilizes the background subtraction operation to calculate the de-background maps, which can remove background interference and enhance crowd features simultaneously. Meanwhile, considering the coarse de-background maps generated by one background removal process, we further cascade multiple CEDMs in PCEDNet to progressively remove background noise and refine the de-background map using a coarse-to-fine strategy. Extensive experiments on five benchmark datasets verify that our PCEDNet achieves superior performance compared with other state-of-the-art approaches. The complexity and time consumption analysis further indicates the lower complexity and computational cost of our network. Code is available at https://github.com/maohule/PCEDNet.