Fast Multiclass Vehicle Detection in Very High Resolution Aerial Images

Abstract

Airborne camera can provide optical images covering a large area at low cost. The collection of traffic and parking data from these images is used for traffic management and monitoring disaster area. In order to retrieve the traffic the vehicles have to be detected and tracked for measuring their speed. The main challenges for such kind of problem are multi-direction detection and multi-type classification. In this thesis, the problem of rapidly detecting multi-class vehicles in aerial images is solved using two stages: the first stage detects vehicles in different directions, and the second estimates the orientations and classify the types of the vehicles. In the first stage, multiple classifers are aggregated to detect different-oriented vehicles. Each binary classifer is trained by AdaBoost algorithm using integral channel features. AdaBoost algorithm combines multiple weak classifers into one strong classifer. Integral channel feature can provide rich feature information as it is the generalization of Haar-like feature constructed on top of the feature channels, which are computed using linear and non-linear transformations of the input image. A soft cascade structure in each binary classifer is used to reject false positives quickly in order to achieve high detection speed. The second stage consists of two artificial neural networks, one is for orientation estimation and the other for vehicle type classification. After the former artificial neural network is trained by multiple classes with different directions, a new input sample can be classifed into corresponding class to realize orientation estimation. Type classification can classify vehicles types, e.g. car or truck. Both artificial neural networks are trained using histogram oriented gradient features from samples. The experimental validation shows that the proposed solution outperforms the baseline in detection accuracy. Moreover, the detection speed is approximately 100 times faster than the brute-force method used in the baseline system.