TraViQuA: Natural Language Driven Traffic Video Querying Using Deep Learning

Video cameras are widely utilized and have ingrained themselves into many aspects of our daily life. Analysis of video contents is more challenging as the size of the data collected from the cameras increases. The fundamental cause of this challenge is because certain data, like the videos, cannot be queried. Our research focuses on converting traffic videos into a structure that can be queried. Specifically, an application called TraViQuA was suggested f or natural language-based car search and localization in traffic videos. To query and identify cars, data including color, brand, and appearance time are used as features. The query is initiated in real time on live traffic feed, as the user enters the search term on the application interface. Our text to SQL conversion algorithm enables the mapping of a search term into a SQL query. Based on the response to the natural language query, TraViQuA can start the video from the relevant time. Deep neural networks were employed in our application for text to SQL conversion and feature extraction. Our research reveals that color and brand models had mean average precision of 98.714% and 91.742%, respectively. The text to SQL conversion had an 80% accuracy rate. To the best of our knowledge, TraViQuA is the first application that enables police officers to input a natural language description of a car and discover the car of interest that matches the description, bridging the gap in traffic video surveillance. Moreover, TraViQuA can be incorporated into other intelligent transportation systems to support law enforcement officials in urgent situations like hit-and-run incidents and amber alerts.