Human Crawl vs Animal Movement and Person with Object Classifications Using CNN for Side-view Images from Camera

An optical camera-based intrusion classification system (Light Intrusion DeTection systEm named as acronym LITE) for an outdoor setting was recently developed by a superset of the authors. The system classified between human and animal images captured in a side-view manner based on the height. Based on the system and algorithm design, most probably human-crawl would be classified as animal by the LITE. In this paper, classification between human-crawl and animal is addressed. In addition to this work, classification of person with weapon versus person with vehicle is also addressed (referred as person with object) to provide more information about the type of intrusions. A Convolutional Neural Network (CNN) based approach is used to solve the above stated two problems. In the case of `` person with object'' classification, a study of different CNN architectures was carried out and analysis corresponding to that is presented. In case of human crawl vs animal movement, performance results corresponding to only the best architecture model is provided among the many tried models. Further on, additional insights are provided about the classification using the attention heat maps and t-SNE plots. The test classification accuracies for human-crawl vs animal and person with object classification on the recorded data are close to 95:65% and 90%, respectively. The LITE, having the Odroid C2 (OC2) Single-Board Computer (SBC) with CNN-based classification algorithm for human-crawl versus animal task ported on it, was deployed in an outdoor setting for a realtime deployment. It provided a classification accuracy close to 92%