Low-Complexity Driving Event Detection from Side Information of a 3D Video Encoder

Mobile phones are often found in cars, for instance when they are used as navigation assistants. This work propose to use their camera, which is often already pointed to the road, to perform some low-complexity analysis of the driving context, with the final aim to detect potentially unsafe conditions. Since content understanding algorithms are typically too complex to run in real time on a mobile device, a driving event detection algorithm is presented based on the side information available from video encoders, which are a highly optimized application in mobile phones. A set of interesting and easy-to-extract features has been identified in the side information and then further reduced and adapted to the specific events of interest. A detection algorithm based on support vector machines has been designed and trained on several hours of video annotated by a human operator to extract the events of interest. The detection algorithm is shown to achieve a good identification rate for the considered events and feature sets. Moreover, results also show that the use of a stereoscopic camera significantly improves the performance of the detection algorithm in most cases.