A TIME-DELAY NEURAL NETWORK ALGORITHM FOR REAL-TIME PEDESTRIAN RECOGNITION
Christian Wöhler, Joachim K Anlaug, Till Portner, Uwe Franke · 1998
In this paper we present an algorithm for recognizing walking pedestrians in sequences of grayscale stereo images taken from a moving camera pair. The method has been designed for use in a driver assistance system for the inner city environment warning the car driver of traffic participants that might cause dangerous situations. Our algorithm is divided into two parts: First, a preliminary detection and tracking stage consisting of a real-time stereo algorithm yields image regions possibly containing a pedestrian. During the subsequent classification stage, these temporal sequences of regions of interest are classified by a feed-forward time delay neural network (TDNN) with spatio-temporal receptive fields. It is possible to stabilize the recognition process byintegrating feedback loops into the TDNN architecture. The complete detection and recognition algorithm runs at a speed of about 70 ms per cycle on a Power PC 604e.