Real-Time Driver Activity Recognition with Random Forests

In this work, we introduce a real-time driver activity recognition method which takes a sequence of depth images as input and outputs an activity class among a predetermined set of driver activities. A classification algorithm called Random Forests is employed and further enhanced by a unique state based inference system to reduce initial classifier errors. For example, frequent changes in driver activities are penalized so as to stabilize the output. The cost of activity change is decided by a state inference system which takes both temporal and spatial coherence into account. The paper will introduce the training system, explain the state inference system and the cost based penalty calculation. Finally we will discuss the results and future work.