A Differential Evolution Based Multiclass Vehicle Detector and Classifier for Urban Environments

Video analytics is emerging as a high potential area supplementing intelligent transportation systems (ITSs) with wide ranging applications from traffic flow analysis to surveillance. Object detection and classification, as a sub part of a video analytical system, could potentially help transportation agencies to analyze and respond to traffic incidents in real time, plan for possible future cascading events, or use the classification data to design better roads. This work presents a specialized vehicle classification system for urban environments. The system is targeted at the analysis of vehicles, especially trucks, in urban two lane traffic, to empower local transportation agencies to decide on the road width and thickness. We present a hybrid appearance model specifically designed for speedy foreground extraction in the given context. We use a simple motion cue based tracking algorithm, and stay clear of using probabilistic trackers. The main thrust is on the accurate detection and classification of the detected objects using an evolutionary algorithm. The detector is backed by a differential evolution (DE) based discrete parameter optimizer. We show that, though employing DE proves expensive in terms of computational cycles, it measurably improves the accuracy of the detection and hence the classification system. The system was tested on multiple real video footage during varied weather conditions from a camera mounted in urban areas achieving a peak classification accuracy of approximately 90%.