Traffic Intersection Vehicle Movement Counts with Temporal and Visual Similarity based Re-Identification

Vehicle movement counting and classification is one of the critical components for traffic intersection monitoring and management. Cameras can be used to determine the vehicle movement counts (left-turning, right-turning, and through movements). Typically, cameras are installed with the view focused towards a particular approach and there is no overlap or very low overlap between the different camera views. Therefore, vehicles need to be re-identified across multiple cameras to detect the complete movement trajectory of the vehicle. In this study, we proposed combining visual similarity obtained using Convolutional Neural Networks (CNN) and temporal similarity (vehicle re-appearance time in cameras) using logistic regression (LR) model to perform vehicle re-identification. The logistic regression model has been used in two stages (without and with hard-negative mining) combining visual and temporal similarity. The results showed that using the hard-negative mining based LR model, the Top@1 results improved by 22% and Top@5 results improved by 8.48%, compared to the results obtained using only visual similarity measure for generating the rankings.