New approach to automatically collect good samples to train a vehicle image-classifier
Chang-Yon Kim, SeungJong Noh, Hae-Rim Shin, Moongu Jeon · 2015
In traffic monitoring systems, it is very important to train an accurate vehicle image-classifier to implement automated video analysis techniques such as detection and tracking. In general, classifiers are obtained from manually collected and labeled training sample images. However, this approach has a problem that it requires significant human efforts to construct dataset. To remedy this drawback, we present a novel method to automatically collect samples, where good samples providing appearance information of vehicles are obtained based on results of background subtraction. Experimental results conducted under highway traffic environments demonstrate effectiveness of the proposed sample collection approach.