Approximately 1300 casualties are caused by blind spot accidents every year in the European Union. Bicyclist and pedestrians are the most common victims.
The most widely used solution is the use of blind spot mirrors, which is obliged in the EU by law since 2003. However, it is shown that these mirrors are often not, or incorrectly used.
We believe that an active driver-independent system offers a better solution. Therefore, our goal is to develop an application that can automatically detect vulnerable road users in blind spot cameras. When a dangerous situation occurs, the system should warn the driver. This is an extremely challenging task.
Firstly, vulnerable road users are a heterogeneous object class. Besides pedestrians, we also need to detect bicyclists, children, wheelchair users and mopeds. Secondly, because the field of view of the camera covers the blind spot area on the side of the truck, we have a highly dynamic background.
Since the camera is moving techniques like adaptive back- ground estimation or background subtraction, which can be calculated fast, are not an option. The biggest challenge, is the hard real-time character of the appli- cation, combined with the need for a high precision and recall rate. We only have limited time available to detect the vulnerable road users.
This paper presents work on a first part of the application: we developed a pedestrian tracker for a moving camera which is both robust and fast. We present a real-time pedestrian tracker that combines a robust appearance-based pedestrian detector with application-specific constraints and motion information.
The targeted application is the automatic detection of vulnerable road users in blind spot cameras on trucks. This application imposes several challenges that need to be tackled. Vulnerable road users are a very diverse class, and we need a high precision and recall rate with real-time performance.
Here, we present a first step towards such an automatic detection system. The novelty of our approach is the extension of a robust pedestrian detector towards real-time performance. The information from the appearance-based detector is used in combination with motion-based estimations to efficiently reduce the search space for the appearance-based detector in consecutive frames.
This results in a multi-pedestrian tracker from a moving camera which is both optimized in terms of accuracy and speed. We recorded several data sequences to evaluate our pedestrian tracker, and performed initial experiments with promising results.
To read this external document in full, download the paper from the author archives online at the KU Leuven Association.