Intelligent Vehicle Counting for Video Surveillance Application

Abstract

In most developed cities in the Philippines, video surveillance systems are presently implemented, which results in vast amounts of video data stored in servers. In a particular city, most CCTV cameras capture traffic scenes in urban roads. Moreover, command centers and traffic management office need this information to respond to emergencies and get visual information on traffic situations. CCTV operators monitor these live-feed data to record and report any incidents that have happened 24/7. With this, little or no time is given for video analysis, such as counting vehicles, which is a piece of essential information in planning and making decisions for the betterment of urban traffic flow and readiness to emergency events. It is for this reason that the researchers proposed an intelligent vehicle counting to be integrated into video surveillance applications using Tensorflow object detection and counting APIs, a deep-learning approach to classify and count vehicles. Initially, a classifier has been built and trained to detect and count vehicles and distinguish them from other objects in the sample captured videos. The newly-trained vehicle classifier is tested and evaluated using recall, precision, and F1-score performance metrics.

On the other hand, the vehicle counter has been run and tested for different positions of the region of interest (ROI) line and eventually evaluated using the accuracy metrics. Results show that the said system is recommended to be integrated into video surveillance systems. However, a fraction of detection and counting errors is noted for future enhancements of the system.