| `` Recent Advances in Computer Vision for Detecting and Counting Objects in Traffic Video'' | |||
| |||
| Abstract: | |||
|
Traffic is an essential aspect of modern society. As people crowd into
huge cities over time, traffic grows to be more difficult to control.
Intelligent surveillance systems have been developed to ease this
growing concern. One of the most common tasks assigned to these systems
is analyzing or retrieving useful information such as vehicle count or
density. Clearly, many approaches have been proposed to tackle the
problem, but computer vision usually remains as an intuitive choice.
Apart from the cheaper and stronger hardware, the reasons computer
vision is chosen mostly come from recent advances in the field itself.
Object counting methods using computer vision can be grouped into two categories: indirect and direct. For indirect approach, locations of object are usually unknown upon counting. Generally, methods follow this approach compromise location information in exchange for speed and counting accuracy. In contrast, direct methods involve finding boundaries or locations of objects. They can be slower but the information they provide are very useful for advanced applications such as tracking or accident detection. Each approach comes with its own strengths and weaknesses, one needs to find the optimal solution to integrate either of them, or both, into their surveillance system. Over the years, a multitude of methods have been proposed in the literature using a wide range of approaches, from texture analysis to deep learning. Recently, state-of-the-art methods have achieved a steep rise in performance that has never been seen before. They will be discussed here as well as the reasons for why they work better and how they are readily to be integrated in real traffic systems. Additionally, insight into remaining problems will also be provided as they are meaningful in shaping future research interests of the community. | |||