Fig. 1. Architecture of DASCNN

Fig. 2. Structure of inception

Fig. 3. Original images and visual density maps. (a) Image 1; (b) density map of image 1; (c) image 2; (d) density map of image 2; (e) color-density scale

Fig. 4. Original images and estimated crowd density maps. (a) Image 1 with truth value of 36; (b) image 1 with estimation value of 31.5; (c) image 2 with truth value of 22; (d) image 2 with estimation value of 21.7

Fig. 5. Comparison of density maps obtained by single-row network and combined network. (a) Image 1; (b) density map of image 1 by shallow network; (c) density map of image 1 by deep network; (d) density map of image 1 by DASCNN; (e) image 2; (f) density map of image 2 by shallow network; (g) density map of image 2 by deep network; (h) density map of image 2 by DASCNN

Fig. 6. Contrast of crowd density estimations. (a) Image 1; (b) density map predicted in Ref. [9]; (c) density map predicted by proposed method; (d) image 2; (e) density map predicted in Ref. [9]; (f) density map predicted by proposed method

Table 1. Dataset information of CAUC-CROWD、UCF-CROWD、AHU-CROWD

Table 2. Comparison of results by proposed method and other algorithms

Table 3. Comparison of results by proposed method and other algorithms

Table 4. Experimental results of estimating number of people from AHU-CROWD