Described is a system for real-time object recognition. The system extracts a candidate target region representing a candidate object from an input image of a scene based on agglomeration of channel features. The candidate target region is classified using a trained convolutional neural network (CNN) classifier, resulting in an initial classified object. A multi-target tracker is used for tracking the classified objects for final classification of each classified object, resulting in a final output, and a device is controlled based on the final output.