Comparison of vehicle detection techniques applied to IP camera video feeds for use in intelligent transport systems

Abstract

Vehicle detection is an important area in Transport and Artificial Intelligence. Through vehicle detection techniques, vehicles can be located across different images. Some of these models are robust enough to identify parts of vehicles in images where the vehicle might be partially occluded. Recent advances in detection methods gave rise to a range of different techniques that can be used for recognition and detection of vehicles. Although each technique has its merits, it is not always the case that the adopted model works well for scenarios involving IP Cameras. The motivation for this study is to compare several state-of-the-art techniques, including deep learning models and computer vision approaches. A set of experiments are developed in order to test these models on a number of low quality IP camera footages set in the transport domain in order to measure detection and recognition accuracy. The final evaluation compares detection accuracy using mean average precision, the semantics of the recognised vehicle as well as recognition robustness when applied to a dataset that contains images with different light conditions. The study also looks at persistence in recognition across frames in video data and a detailed description of the dataset used to train the evaluated models. Finally, the paper also goes through some scenarios that applies the results obtained in this study to ITS systems that use IP camera feeds.