Motion Clustering Estimation on Video Sequences Using Kohonen’s Self Organizing Map (SOM) Neural Networks

Motion estimation is a very important and interesting area of research. It has become the necessity of many fields such as agriculture, security, medicine, traffic, and sports, the growth of a plant, tracking the movement of a vehicle within traffic, or observing the movements of a runner's hands or legs. Traditional methods for motion estimation estimate the motion field between a pair of images as the one that minimizes a predesigned cost function. An unsupervised learning method from the family of artificial neural networks i.e. Kohonen;s Self-organizing Map or SOM, a popular clustering way, based on Euclidian distance, when at test time, is given a pair of images as input it produces a dense motion field as its output layer. In the absence of large datasets with ground truth motion that would allow classical supervised training methods, the network in an unsupervised manner using the Self-organizing map is used for the training and hence for clustering to find the dense active regions in multiple frames of a video sequence. SOM used in this paper is also compared with other methods of clustering like k-Means algorithm, Nearest neighbour algorithm, Image subdivision algorithm, and Competitive learning network. Consequently It is observed that the Self-organizing map provides more accurate results and less error.