Pedestrian detection using two-stage sparse coding algorithm

Detecting pedestrians efficiently and accurately is the first fundamental step in Intelligent Transformation Systems. In this paper, a novel and simple pedestrian detector with two stage of feature extraction and supervised classifier was proposed. The main contribution of the system is composed of two parts: (1) proposing a two-stage of feature extraction in Pedestrian detection system (PDS) while previous works only containing one stage. And the second-stage is fed with the output of the first stage. The systems with two stages of feature extraction can receive better accuracy than one. Generally, feature exaction stages include a filter bank, a non-linear transformation and some sort of feature pooling layer. (2) We extract feature in an unsupervised fashion. The filter in the feature extractor stages are initialized using sparse coding algorithm. A sparse coding algorithm can learn representations that are not only sparse, but also invariant to some transformation which is suit for pedestrians in real environment. Also sparse coding algorithm can use in real-time object recognition such as pedestrian detection. Experiments show the proposed system can achieve both high detection rate.