Deep neural network based object detection and classification in drone and cameratrap images

Drones and ground level camera traps are among the most popular image acquisition technologies used for ecological applications today. They can capture image data from perspectives and environments typically difficult for humans to access. Their advantages include low-cost operation, long working times, greater time-saving compared to human efforts, and improved efficiency of capturing effective images.

Drone and camera trap captured image data offer rich information for computer vision applications enabling automated image analysis. However, existing object detectors, based on traditional computer vision and/or machine learning algorithms and more recent deep neural network based models, face significant challenges in detecting and classifying target objects in images captured from drones and camera traps. This is due to the distinct appearance and characteristics of objects in these images as compared to those obtained through traditional methods. Firstly, drones typically operate at high flight altitudes, and hence the size of objects is significantly reduced. Camera traps, conversely, may capture distant targets when the sensor is triggered. Secondly, the environments in which drones and camera traps operate result in unavoidable variations in lighting, capture angles, backgrounds, and overlapping or blurred images due to likely target movements. This is true, especially for camera traps that operate both day and night. Finally, some targets are challenging to recognize and classify even for humans, limited by the lens focusing capabilities and photographic clarity of modern cameras. These aspects increase the challenge of identifying and classifying targets automatically using computer vision.