Video-based surveillance pedestrian detection is playing a key role in emerging technologies, such as Internet of Things and Big Data for use in smart industries and cities. In pedestrian detection, factors, such as lighting, object collisions, backgrounds, clothes, and occlusion cause complications because of inconsistent classification. To address these problems, enhancements in feature extraction are required. These features should arise from multiple variations of pedestrians. Well-known features used for pedestrian detection involve histogram of gradients, scale-invariant feature transform, and Haar built to represent boundary level classifications. Occlusion feature extraction supports identification of regions involving pedestrian detection. Classifiers, such as support vector machine and random forests are also used to classify pedestrians. All these feature extraction and pedestrian detection methods are now being automated using deep learning methods known as convolutional neural networks (CNNs). A model is trained by providing positive and negative image data sets, and larger data sets provide more accurate results when a CNN-based approach is used. Additionally, Extensible Markup Language cascading is used for detecting faces from detected pedestrian.