Lightweight convolutional neural network for vehicle recognition in thermal infrared images

Nighttime vehicle recognition based on driver assistance systems (DAS), targeted at enhancing the driver perception of the environment, has been of interest to researchers in various disciplines. Research to date has primarily focused on the development of automatic, low-power, and small systems. Although popular in the strategic sector, thermal infrared cameras have not been considerably explored in traffic. This study demonstrates the use of thermal infrared cameras in the field of DAS and discusses their performance in nighttime vehicle recognition. In this work, a dataset of thermal infrared images comprising four vehicle classes was generated: bus, truck, van, and car. To achieve robust vehicle recognition, we designed a convolution neural network (CNN) called Net1 having four convolution layers for classifying four vehicle classes. To achieve compact and efficient performance, this paper applied the Fire module which was the core structure of SqueezeNet to build nine lightweight CNNs (Net2-Net10). Using Net1 as the reference network, we compared the performance of the above nine networks. Amongst all the networks, Net9 was found to be the optimal one with a 97% classification accuracy with 10.6% of the number of parameters of Net1. The recognition time of a single picture was only 0.52 ms on a desktop with a common CPU, thereby enabling its deployment in mobile terminal systems and embedded devices like DAS.