Traffic Scene Analysis and Classification using Deep Learning

In this study, we aim to use new deep-learning tools and convolutional neural networks for traffic analysis. ResNeXt architecture, one of the most potent architectures and has attracted much attention in various fields, has been proposed to examine the scene, and classify it into three categories: cars, bikes (bicycles/motorcycles), and pedestrians. Previous studies have focused more on one type of classification and reported only human-facial recognition or vehicle detection. In contrast, the proposed method uses precise architecture to perform the classification of three classes. The proposed plan has been implemented in several steps: the first stage is to divide the critical objects. In the next step, the characteristics of the obtained objects are extracted to classify the process into three classes. Experiments have been conducted on different and essential datasets such as high-traffic, low-quality, real-time scenes. Essential evaluation criteria such as accuracy, sensitivity, and specificity show that the performance of the proposed method has improved compared to the methods being compared. The accuracy criterion reached more than ۹۲%, sensitivity about ۸۹%, and specially to ۹۰.۲۵%. The proposed method can be used to implement intelligent cities, public safety, and metropolitan decisions and use the results in urban management, predictive modeling of lost data management, sequential data management, and generalizability.

کلیدواژه ها:

Classification ، convolutional neural network ، Deep Learning ، ResNeXt ، Traffic scene analysis