Sensor networks are appealing because they can be used to monitor real-life environmental conditions. One such application is the behavioral analysis of birds. The challenge in real-life monitoring is having an end-to-end system so that the tedious task of analyzing large data can be avoided. This paper shows that it is indeed possible to have a complete system with high accuracy.

The basic components needed to build a complete vision-based avian monitoring system are: an imaging system that works in low-light conditions, and the ability to detect the presence or absence of birds, to count eggs, and to gather summary statistics for the identification of pre-incubation, incubation, and hatching cycles. An infrared light-emitting diode (LED) is used to enhance imaging in low-light conditions. The bird detection algorithm is based on interest point counting--that is, counting the regions where the image exhibits large gradients in two independent directions. Since it also assumes that birds will have smooth feathers, the interest point count is less during a bird’s presence. Egg counting is the most difficult part in the inference process, as it is a multiclass decision process. The usual approach of multiple object detection in clutter fails in egg counting because eggs are generally of uniform intensity. Instead, the detectors often classify surrounding nesting materials. Therefore, the traditional algorithm is augmented by heuristics, such as the temporal association of images to enforce increments of, at most, one per day during pre-incubation, no change during incubation, and rapid changes during hatching period. The nesting stage is determined using the egg counts.

This is a great attempt to automate the complete system for environmental monitoring. However, it is still a prototype, as the research is performed in a controlled environment. The imaging will be much more challenging for birds that make their own nests, and the bird detection might not be suitable for birds that have contrasting feathers. Furthermore, the egg count might have to be performed without temporal heuristics, as some birds lay eggs more than once a day and other birds might be tempted to use the same nest. Whatever the complexities may be, this research will surely be considered as a basis for future work on vision-based monitoring systems.