Unsupervised pedestrian sample extraction for model training

Many researches on pedestrian detection use benchmarking datasets such as INRIA for model training. However, models trained with standard video database do not usually obtain satisfying performance in real-life conditions. Hence, supervised training through manually labelled instances is often required to help achieving better detection result. In this research, an innovative unsupervised training approach is proposed. By analyzing the histogram of adjacent pixels modelled from the video sequences, separated pedestrians can be extracted without manual intervention. Experiments have shown consistent performance that is superior over the state-of-the-art methods.