ARNet: Accurate and Real-Time Network for Crowd Counting
| Autor: | Wenyue Wei, Yinfeng Xia, Qing He, Baoqun Yin |
|---|---|
| Rok vydání: | 2021 |
| Předmět: | |
| Zdroj: | PRICAI 2021: Trends in Artificial Intelligence ISBN: 9783030893620 PRICAI (2) |
| Popis: | Taking into account the problem of the redundant structure and excessive parameters in the current crowd counting network, we propose an end-to-end encoder-decoder architecture called Accurate and Real-time Network (ARNet) for high-accuracy and real-time crowd counting. The encoder adopts lightweight SqueezeNet as the backbone network to extract multi-level features, the decoder can integrate contextual information to enhance the semantic representation capabilities of low-level features. In addition, we design the Parameter-Sharing Context-Aware Module (PSCAM) and the Mask Density Generator (MDG). Without adding excessive parameters, the PSCAM can capture the global context by applying multiple dilated convolutional layers with the same convolution parameters and different dilation rates. The MDG based on multi-task learning can generate accurate density maps by introducing semantic segmentation to suppress background interference. Extensive experiments on four benchmark crowd datasets, which indicate our ARNet can achieve the optimal trade-off between counting accuracy and computation efficiency. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|