Traffic Surveillance: Vehicle Detection and Pose Estimation Based on Deep Learning

Video-based traffic surveillance analysis is an important area of research with numerous applications in intelligent transportation systems. Due to camera positioning, background crowd, and vehicle orientation fluctuations, urban situations are more complex than highways. This paper provides a state-of-the-art technique for vehicle detection and orientation estimation based on the convolutional neural network CNN for detecting and determining the orientation of a vehicle from a given image to reduce traffic accidents. Different CNN model architectures have been examined to reach this approach's goal, which results in a small and fast model that is compatible with limited-resources hardware. A large-scale dataset of vehicles has been used to train the model. The dataset includes different types and views of cars; the taken images are high quality with diverse backgrounds and light conditions. To train the model, the dataset has been divided into five classes according to view: Front, Rear, Side, Front-side, and Rear-side, to fit the requirement of this work. The system achieves a high accuracy result.