Feature selection and neural network architecture evaluation for real-time video object classification

The convergence of the proliferation of inexpensive high quality video cameras, the current global security-aware environment, and increasing capability of newer generation computer hardware enables the use of computational resources to aid human operators in the video surveillance task by not only alerting on scene change detection, but providing alerts based on object classification. Current techniques used in real-time classification of video objects tend to use fast classifiers, such as linear support vector machines and decision trees due to the historic limitation of computational resources. Furthermore, the features fed to these classifiers tend to be small in size or sparse, again, due to limited computational resources. The new computational devices, which are now readily available, provide the resources to enable non-linear classification techniques, such as complex neural networks operating on large dense feature vectors, and convolutional neural networks. In this paper, we compare and evaluate several neural network architectures and their corresponding features that are capable of classifying several video objects in real-time against the Pascal 2007 dataset.