3D vision sensing for improved pedestrain safety

Abstract

Pedestrian-vehicle accidents account for the second largest source of automotive related fatality and injury worldwide. Automotive manufactures will soon be required to meet regulations specifying safety requirements for pedestrian-vehicle collisions. The inclusion of pedestrian protection systems (eg. External airbags) is being consider as a solution to preventing pedestrian fatality and injury. However, such systems require knowledge of pedestrian presence for correct activation. This thesis describes work towards a computer vision system to detect pedestrians which could fulfil the sensory requirements for activating automotive pedestrian protection devices.

In this work, the requirements for a pedestrian sensor were examined and a prototype vision system was developed to demonstrate the concepts discussed in the thesis. To achieve greater robustness and an improved understanding of the environment, we focussed on using 3D and temporal techniques combined with existing pedestrian detection methods. ¶ Stereo vision was employed to provide 3D information about the scene. The well known computer vision concept of disparity maps was used to generate a 3D scene representation. Additional vision algorithms were developed to provide scene understanding and thus segment a scene into obstacles (pedestrians, vehicles and other road infrastructure). Two methods were investigated for this purpose: Inverse Perspective Mapping and v-disparity, with the latter producing superior results, and thus v-disparity was used for 3D obstacle segmentation. ¶ Next, we focused on developing a method to classify detected obstacles as either pedestrian or non-pedestrian. Existing algorithms which examine a pedestrian’s shape and provide a classification result using Support Vector Machines were used to fulfil this obstacle classification task. We extended the existing work to include a pedestrian model from a front/rear and side poses. ¶ Finally, temporal information from both the obstacle detection and classification results were used to enhance system results. We used Kalman filtering techniques to track pedestrians and provide motion predictions. Additionally, Bayesian probability was used to provide a certainty of pedestrian detection based on an object’s classification history. This provided greater robustness to the overall detection results. ¶ The developed prototype was installed on two vehicles, a Toyota Landcruiser and a Volvo S80, to perform real world testing. Results from the prototype were excellent, achieving average detection rates of 83% with average false detection rates of only 0.4%. Show more