This paper presents a method to generate a dataset for training a deep convolutional network to detect a non cooperative unmanned aerial vehicle in video data. Deep convolutional network have shown a great potential for tasks like object detection and have been continuously improved in the last years. Still, the amount of training data is large and their generation can be complex and time consuming, especially if the appearance of the detected object is not clearly specified. The concept presented here is to train a deep convolutional neural network just with a few two dimensional images of unmanned aerial vehicle to simplify the process of generating training data. Performance of the trained network is evaluated with data from real experimental flights and compared with hand-labeled ground truth data to validate the correctness. To cover situations when the classifier fails at the detection, the output is integrated in a image processing pipeline for object tracking in order to establish a continuous tracking.