Multiple Kernel Learning for vehicle detection in wide area motion imagery

Vehicle detection in wide area motion imagery (WAMI) is an important problem in computer science, which if solved, supports urban traffic management, emergency responder routing, and accident discovery. Due to large amount of camera motion, the small number of pixels on target objects, and the low frame rate of the WAMI data, vehicle detection is much more challenging than the task in traditional video imagery. Since the object in wide area imagery covers a few pixels, feature information of shape, texture, and appearance information are limited for vehicle detection and classification performance. Histogram of Gradients (HOG) and Haar descriptors have been used in human and face detection successfully, only using the intensity of an image, and HOG and Haar descriptors have different advantages. In this paper, we propose a classification scheme which combines HOG and Haar descriptors by using Generalized Multiple Kernel Learning (GMKL) that can learn the trade-off between HOG and Haar descriptors by constructing an optimal kernel with many base kernels. Due to the large number of Haar features, we first use a cascade of boosting classifier which is a variant of Gentle AdaBoost and has the ability to do feature selection to select a small number of features from a huge feature set. Then, we combine the HOG descriptors and the selected Haar features and use GMKL to train the final classifier. In our experiments, we evaluate the performance of HOG+Haar with GMKL, HOG with GMKL, Haar with GMKL, and also the cascaded boosting classifier on Columbus Large Image Format (CLIF) dataset. Experimental results show that the fusion of the HOG+Haar with GMKL outperforms the other three classification schemes.