Improve Crowd Size Estimation by Leveraging Deformable Convolutional Neural Network and Deformable Region of Interest
Lubamba Kasangu Eric, Kevang Cheng, Rabia Tahir, Mohammad Saddam Khokhar · 2019
Either in still image or sequence images (video), counting is a challenging task due to many factors such as: scale variations, crowded scene, lighting, orientation, camera position and pedestrian appearance. Despite the fact that practitioners have applied deep network to deal with crowd counting and its increase gain in momentum, challenges still remain. Therefore, we propose a new model for Crowd Estimation with help of Deformable Convolutional Neural Network (DCN) and Deformable Region of Interest (DRoI) pooling. The proposed model is mainly divided in two blocks: a front-end for feature extraction and a back-end comprised by larger reception fields. In both blocks, we replaced standard pooling with Deformable Region of Interest Pooling. The experiment results show better accuracy performance, cost effectiveness of the network and robustness of our model, for less parameters are employed in the proposed model during training as compared to previous models.