Human Detection using Illumination Invariant Feature Extraction for Natural Scenes in Big Data Video Frames

This research proposes a reliable machine learning based computational solution for human detection. The proposed model is specifically applicable for illumination-variant natural scenes in big data video frames. In order to solve the illumination variation problem, a new feature set is formed by extracting features using histogram of gradients (HoG) and linear phase quantization (LPQ) techniques, which are combined to form a single feature-set to describe features in illumination variant natural scenes. Pre-processing is applied to reduce the search space and improve results, and as the humans are in constant motion in the frames, a search space pruning algorithm is applied to reduce the search space and improve detection accuracy. Non-maximal suppression is also applied for improved performance. A Support Vector Machine (SVM) based classifier is used for fast and accurate detection. Most of the current state-of-the-art detectors face numerous problems including false, missed, and inaccurate detections. The proposed detector model shows good performance, which was validated using relevant UCF and CDW test data-sets. In order to compare the performance of the proposed methodology with the state-of-the-art detectors, some selected detected frames were chosen considering their Receiver Operating Characteristic (ROC) curves. These curves are plotted to compare and evaluate the results based on miss rates and true positives rates. The results show the proposed model achieves best results.