Improving the Accuracy of Pedestrian Detection in Partially Occluded or Obstructed Scenarios

More than 90% of traffic accidents are caused by humans. Therefore, the wide adoption of autonomous vehicles and advanced driver assistance systems (ADAS) is expected to save hundreds of thousands of lives in the next decades. One important capability of autonomous vehicles and ADAS is pedestrian detection. However, such detection becomes very challenging in scenarios where pedestrians are partially occluded, resulting in high rates of non-detection. In this paper, we improve the performance of a pedestrian classifier by proposing frameworks composed of Histogram of Oriented Gradients (HOG) feature extraction, combined with Support Vector Machine (SVM), and eXtreme Gradient Boosting (XGBoost) learning models. In order to validate the performance of the improved approach, we consider the PSU and the INRIA pedestrian datasets. The state-of-the-art detection rates for the PSU and INRIA datasets are 55% and 54%, respectively. The proposed approach achieves detection rates of 86% and 82%, respectively, considerably outperforming the state-of-the-art results.