Human-Crowd Density Estimation Based on Gabor Filter and Cell Division

Human-crowd density estimation problem has always been difficult when the scenario is affected by strong perspective distortion and high occlusion. However, this difficulty can be mitigated by the indirect counting approach, i.e. counting them without actually detecting them. Based on this approach, Qing Wen et al. proposed a method relies on the texture features extraction using Gabor filters and least square support vector machine. Our proposed method, inspired by their algorithm, uses semi background masking to eliminate redundant areas of filtered image. The local texture features are extracted after applying grid which divides the image into fixed cells. Experiments are done on the same dataset used in Qing Wen´s work, the PETS2009, for a better comparison. The results have presented that our proposed method is more accurate and efficient.