3-D Colour-based Segmentation and Recognition for Environmental Monitoring

This work focuses on the automatic monitoring of the natural environment using images of animals. An unsupervised 3D colour texture segmentation and recognition algorithm is developed and applied in the recognition of natural patterns (monkeys) in outdoor image scenes (Amazon Forest). We present an unsupervised segmentation algorithm, which uses 3D Graph-theoretical clustering for colour images. This approach is based on image chromaticity and intensity. Subsequently each cluster of the segmented image is represented by six colour texture angle indices [1] [2] and the Frobenius norm is applied to measure the distance between candidate and prototype colour images. The robustness of this algorithm to viewpoint transformations (rigid, scaling, perspective...), change of illumination (intensity, spectral power distribution...), non-rigid change of shape and background and the suitability to operate in a co-operative integrated system are considered very important characteristics. We demonstrate the segmentation approach using colour images of natural scenes from the Amazon Forest and we provide experimental results that compare the recognition algorithm with competing methods using colour invariants. The effect of the segmentation on the performance of the recognition algorithm is also considered.