Classification of Birds/Drones Tracks using Images Kinematics and Texture Filters

Abstract This work presents techniques for the classification of drones and birds based on videos of those objects while flying. With the unavailability of radars Doppler and micro-Doppler signals and instead of relying on static images of the drones or birds, this work proposes methods for calculating the kinematics features and texture features extracted by the Gabor filter and GLCM filter to use them in classifying the tracks generated from the videos. Machine learning techniques such as SVM, Random Forest, Shallow Neural Networks, and CNN (Deep Neural Networks) are used in the classification. Open-source data sets of tracks of birds and drones are utilized. Comparisons are made between the different methods. Hybrids of several features used in the classification proved to have better results than using one type of feature. The deep NN (CNN), which is fed images directly, did not beat the SVM, RF, and shallow NN with the hand-crafted features used as inputs. The testing showed that the best results are achieved using combinations composed of kinematics and Gabor or GLCM extracted features. However, based on our work outcome, Random Forest (RF) with hybrid features showed the best results. Its performance had the highest values of precision and accuracy of 95% and 91% respectively, as opposed to SVM and the deep NN methods. We believe, there is a need for a larger and more diverse set of training and testing data to improve performance.