A Lightweight Moving Vehicle Classification System Through Attention-Based Method and Deep Learning

The convolutional neural network (CNN) has shown excellent benefits in the classification of objects in the latest years. An important job in the context of intelligent transportation is to properly identify and classify vehicles from videos into various kinds (e.g., car, truck, bus, etc.). For monitoring, tracking and counting purposes, the classified vehicles can be further evaluated. At least two major difficulties stay, however; excluding the uninteresting area (e.g., swinging movement, noise, etc.) and designing an effective and precise system. In order to obviously differentiate the interesting region (moving car) from the un-interesting region (the rest of the area), we introduce a novel attention-based approach. Finally, to significantly increase the classification efficiency, we feed the deep CNN with the respective interesting region. We use several challenging outdoor sequences from the CDNET 2014 (baseline, bad weather and camera jitter classes), and our own dataset to assess the proposed approach. Experimental results show that it costs around ~85 fps in GPU (and ~50 fps in CPU) to classify moving vehicles and maintaining a highly accurate rate. Compared with other state-of-the-art object detection approaches, our method obtains a competitive detection accuracy. In addition, we also verify the result of the proposed approach by comparing with recent 3D CNN method, called saliency tubes.