Abstract
Scale variation because of perspective distortion is still a challenge for crowd analysis. To address this problem, an atrous convolutions spatial pyramid network (ACSPNet) is proposed to perform crowd counts and density maps for both sparse and congested scenarios. Atrous Convolutions sequenced with increasing atrous rates are utilized to exaggerate the receptive field and maintain the resolution of extracted features. Different rates of atrous convolution blocks in the pyramid are skip-connected to integrate multi-scale information and extent scale perception ability. Atrous Spatial Pyramid Pooling (ASPP) is employed to resample information at different scales and contain global context. We evaluate our ACSPNet on five challenging benchmark crowd counting datasets and our method achieves state-of-the-art mean absolute error (MAE) and mean squared error (MSE) performances.
Original language | English |
---|---|
Pages (from-to) | 91-101 |
Number of pages | 11 |
Journal | Neurocomputing |
Volume | 350 |
DOIs | |
Publication status | Published - 20 Jul 2019 |
Keywords
- Atrous convolutions
- Crowd counting
- Crowd density estimation
- Multi-scale