Crowd counting is an important task that shown great application value in public safety-related fields, which has attracted increasing attention in recent years. In the current research, the accuracy of counting numbers and crowd density estimation are the main concerns. Although the emergence of deep learning has greatly promoted the development of this field, crowd counting under cluttered background is still a serious challenge. In order to solve this problem, we propose a ScaleAware Crowd Counting Network (SACCN) with regional and semantic attentions. The proposed SACCN distinguishes crowd and background by applying regional and semantic self-attention mechanisms on the shallow layers and deep layers, respectively. Moreover, the asymmetric multi-scale module (AMM) is proposed to deal with the problem of scale diversity, and regional attention based dense connections and skip connections are designed to alleviate the variations on crowd scales. Extensive experimental results on multiple public benchmarks demonstrate that our proposed SACCN achieves satisfied superior performances and outperform most state-of-the-art methods. All codes and pretrained models will be released soon.