IRUVD: a new still-image based dataset for automatic vehicle detection

One of the difficult tasks in the field of computer vision is the classification and detection of vehicles. Researchers from all over the world are working to create autonomous vehicle detection (AVD) systems due to their numerous practical applications, including highway management and surveillance systems. Deep learning techniques, which require a lot of data for proper model training, are the current AVD trend. However, a number of vehicles are discovered in India, the second-largest nation in terms of population, that are not included in the vehicle detection datasets that are currently in use. Furthermore, India's over crowding makes traffic management difficult and unusual. In this research, we present a dataset for still-image-based vehicle detection that includes one class of pedestrians and 13 different types of vehicles that are seen on Indian urban and rural roads. Initially, we provide baseline results using some state-of-the-art deep learning models on this dataset. To improve the accuracy further, we present an ensemble-based object detection and classification model. The dataset consists of 4K images and 14.3K bounding boxes of various vehicles; that is researchers are provided with appropriately annotated rectangular boxes for use with these vehicles in the future. A 16-megapixel Sony IMX519 high-resolution camera was used to take all images while travelling throughout West Bengal, an Indian state on the eastern side. Dataset can be found at:https://github.com/IRUVD/IRUVD.git.

Automated summary

One of the challenging tasks in computer vision is the classification and detection of vehicles. Researchers worldwide are working on autonomous vehicle detection systems, which have various practical applications. The current trend in AVD is deep learning techniques, although many Indian vehicles are not included in the existing detection datasets. In this research, a dataset for still-image-based vehicle detection is presented, including one class of pedestrians and 13 different types of vehicles commonly seen on Indian roads. A baseline result is provided using state-of-the-art deep learning models, and an ensemble-based object detection and classification model is proposed to further improve accuracy. The dataset consists of 4K images and 14.3K bounding boxes, providing researchers with annotated rectangular boxes for future use.