Real-time detection and classification of cars in video sequences

We present a system capable of detecting cars in gray-valued videos of traffic scenes based on easy-to-compute orientation selective features derived from gradient filter outputs. The car detection system consists of two processing stages (initial detection and confirmation) and is embedded into a comprehensive architecture of interacting modules optimized for various aspects of driver assistance applications. The initial detection stage uses a heuristic for generating hypotheses which are then presented to a single neural network (NN) classifier for confirmation, which is trained on examples in a supervised way. We show that one can achieve approximate scale-invariance in the confirmation stage by using approximately scale-invariant image features and training with differently sized examples. .The NN used for confirmation are optimized using a simple pruning algorithm. The dependence of detection accuracy and network complexity is investigated; we find that extremely simple networks give surprisingly good classification accuracies at very high speed.