Feature Synthesization for Real-Time Pedestrian Detection in Urban Environment.

Real-time pedestrian detection is very essential for auto assisted driving system. For improving the accuracy, more and more complicate features are proposed. However, most of them are impracticable for the real-world application because of high computation complexity and memory consumption, especially for onboard embedding system in the unmanned vehicle. In this paper, a novel framework that utilizes reconstruction sparsity to synthesize the feature map online is proposed for real-time pedestrian detection for the early warning system of the unmanned vehicle in real world. In this framework, the feature map is computed by sparse line combination of the representative coefficient and the feature response of trained basis which is learned offline. The efficiency of our method only depends on the dictionary decomposition no matter how complicated the feature is. Moreover, our method is suitable for most of the known complicate features. Experiments on four challenging datasets: Caltech, INRIA, ETH and TUD-Brussels, demonstrate that our proposed method is much efficient (more than 10 times acceleration) than the state-of-the-art approaches with comparable accuracy.