Deep Neural Networks as Feature Extractors for Classification of Vehicles in Aerial Imagery

Detection of vehicles from aerial images have several real world implications in surveillance, military applications, traffic lot management, border patrol and traffic monitoring. The system proposed in this paper intends to automate the process of detecting vehicles from aerial images, rather than relying on a human operator. Here, we identify an optimum classification strategy for the proposed detection system, which is the initial stage of designing a vehicle detection pipeline. This research focuses on the feature extraction capabilities of standard neural network models like, Alexnet [6] and VGG-16 [7], which are compared against classic feature extraction techniques, like Histogram of Oriented Gradients and Singular Value Decomposition. The extracted features are benchmarked across standard machine learning algorithms such as Support Vector Machine and random forest. It is observed that the neural net extracted features gives an overall classification accuracy of 99% on the VEDAI dataset. The classification was treated as a binary class problem with vehicles as one class and rest everything as non-vehicles.