Conclusion

The main contribution of this paper is the novel image segmentation method based on discretization of the color space domain and the adaptive kernel density estimation. For the image segmentation task, spatial constraints of coherent image objects are recovered using the region growing technique where the growing criterion is the affiliation of subsequent pixels to same range domain cluster. Important request for robustness and stability of the segmentation algorithm operating in the erratically changing environment is fulfilled by implementing the variable bandwidth kernel. Using two-step density estimation approach, the kernel bandwidth is made locally adaptive to the data without considerable increase in the computational complexity. Adaptivity to the noise in the input scene is attained using relative noise level. Experimental evaluation of the proposed algorithm has shown that the quality of the provided segmentation is comparable to those of the Mean Shift algorithm, widely adopted in the vision community, while running times of the FHS algorithm are several times faster.

The proposed segmentation technique presents the efficient and versatile low-level tool, which can be easily integrated with other image processing techniques. The FHS method is implemented in the intelligent Forest Fire (iForestFire) video surveillance system [37], within the module for automatic forest fire detection. Output is taken from several stages of the FHS algorithm and integrated with other low-level techniques like color-based pixel classification, edge-preserve filtering and shape analysis.

General grid-based clustering technique proposed in section 4 does not scale well with the dimension of the feature space, due to the time and space complexity of multidimensional mesh processing and the curse of dimensionality problem inherent to all density-based approaches. However, this technique should be applicable to the processing of other low-level image features, as well as to other problems when efficient processing of large number of low-dimensional data samples is required.