Anomaly detection in wearable location trackers for child safety

Abstract

Child Location trackers are widely available devices but the ability of these trackers to detect outlying incidents remain untapped. The main goal of this research is to incorporate anomaly detection in wearable child location trackers using machine learning, with a primary focus on unsupervised outlier detection techniques. Therefore, a low powered gyroscope-enabled electronic tracker using Low Drop Out (LDO) converters was fabricated as a building block for the software/hardware integration. Common unsupervised techniques were adopted for experimentation including the Local Outlier Factor (LOF), the k-Nearest Neighbor (kNN) Global Anomaly Score and the Histogram-based Outlier Score (HBOS). The results obtained show that LOF performs well in detecting local outliers as well as identifying new patterns to prevent false alarms, however, the paths to the newly identified locations bring about false alarms. The kNN Global Outlier Score is only capable of detecting global outliers where k is greater than 15. HBOS, on the other hand, performs better than the kNN Global Outlier Score but is unable to learn new locations that become part of the normal pattern. The proposed model is a potential candidate for an outlier detection system for combating child loss and kidnapping cases, thereby ensuring child safety globally.