Attention-based Approach for Efficient Moving Vehicle Classification

摘要

In recent years, the convolutional neural network (CNN) has shown great advantages in object classification. In the context of smart transportation, an essential task is to correctly detect vehicles from videos and classify them into different types (e.g., car, truck, bus, and etc.). The classified vehicles can be further analyzed for surveillance, monitoring, and counting purposes. However, at least, there are two main challenges remain; excluding the un-interesting region (e.g., swaying motion, noise, etc.) and designing an efficient and accurate system. Therefore, we introduce a novel attention-based approach in order to clearly distinguish the interesting region (moving vehicle) with the un-interesting region (the rest of the region). Finally, we feed the deep CNN with the corresponding interesting region to boost the classification performance considerably. We evaluate our proposed idea using several challenging outdoor sequences from the CDNET 2014 and our own dataset. Experimental results show that it costs around ~85 fps to classify moving vehicles and keep a highly accurate rate. In addition, compared with other state-of-the-art object detection approaches, our method obtains a competitivef-measure score.