Whale Detection
Whale Detection
Characterization of dynamically varying shapes using Neural Networks
We have developed a new methodology based on neural networks to characterize the variation of a shape pattern in monochromatic video where a dynamic texture forms the background. The technique is demonstrated by developing an application that can count whales in the sea by detecting whale blows. The input to the application is an infrared video captured by a camera located at an elevated position on the shore and directed towards the sea. The algorithm is designed based on the spatial and temporal characteristics of whale blows. The first part of the algorithm consists of thresholding techniques that filter out the possible candidates to a group containing whale blows and certain textures on the sea. A novel thresholding technique called grid thresholding is proposed so that the detector is able to detect very small blows while keeping the number of false positives to a minimum. Grid thresholding is adaptive and forms the first part in the algorithm. Two measures called Cumulative Absolute Difference (CAD) and Cumulative Difference (CD) are defined that can differentiate between whale blows and other textures present in the scene. As the final part of the detection algorithm, we have used a neural network to differentiate between whale blows and the various textures on the surface of the sea. We have also developed a preliminary tracking method based on timing of whale blows to track the movement of whales.
