Cost Optimization of Volumetric Surveillance for Sky Monitoring : Towards Flying Object Detection and Positioning

Unlike surface surveillance, volumetric monitoring deals with three-dimensional target space and moving objects within it. In sky monitoring, objects fly within outdoor and often remote volumes, such as wind farms and airport runways. Therefore, multiple cameras should be implemented to monitor these volumes and analyze flying activities. Due to that, challenges in designing and deploying volumetric surveillance systems for these applications arise. These include configuring the multi-camera node placement, coverage, cost, and the system's ability to detect and position flying objects. The research in this dissertation focuses on three aspects to optimize volumetric surveillance systems in sky monitoring applications. First, the node placement and coverage should be considered in accordance with the monitoring constraints. Also, the node architecture should be configured to minimize the design cost and maximize the coverage. Last, the system should detect small flying objects with good accuracy. Placing the multi-camera nodes in a hexagonal pattern while allowing overlap between adjacent nodes optimizes the placement. The inclusion of monitoring constraints like monitoring altitude and detection pixel resolution influences the node design. Furthermore, presented results show that modeling the multi-camera nodes as a cylinder rather than a hemisphere minimizes the cost of each node. The design exploration in this thesis provides a method to minimize the node cost based on defined design constraints. It also maximizes the coverage in terms of the number of square meters per dollar. Surveillance systems for sky monitoring should be able to detect and position flying objects. Therefore, two new annotated datasets were introduced that can be used for developing in-flight birds detection methods. The datasets were collected by Mid Sweden University at two locations in Denmark. A YOLOv4-based model for birds detection in 4k grayscale videos captured in wind farms is developed. The model overcomes the problem of detecting small objects in dynamic background, and it improves detection accuracy through tiling and temporal information incorporation, compared to the standard YOLOv4 and background subtraction