182 / 2016-12-24 06:22:46

Novel Adaptive Learning Scheme for GMM

12215,12216,8274

Gaussian Mixture Modelling (GMM) is widely used background subtraction (BS) method of object detection. It deals robustly with practical issues such as illumination change and multimodal background. Appropriate selection of critical parameters such as learning rate (α) and threshold (T) is the key factor of satisfactory performance of BS. Traditionally, these parameters are selected manually and kept constant over processing of video. Programmer must be expert in appropriate selection else, it is quite difficult to use GMM. Constant parameters couldn’t tackle background varying with random speed and may cause false detection. Therefore, proposed work defines novel adaptive learning scheme which involves automated selection and adaptation of learning rate with fixed threshold T. A new parameter Effective Intensity Change (EIC) is introduced which extract background dynamics at every frame instance. Learning rate is evaluated based on EIC occupancy and applied to adapt background. Special scheme has been developed to deal with sudden illumination change which significantly improves performance of GMM. This approach removes performance dependency of GMM on initial selection of learning rate. Novel adaptive learning scheme has been evaluated over Wallflower dataset. Experimental result shows that this scheme provides remarkable improvement in detection as compared to other tuning methods.