| Popis: |
There are several women safety devices in the market today. Women who suffer these atrocities are even denied basic human rights, as set out in the Criminal Code. Women who are not as fit (physically) as men need to be protected from the evils of society. The introduction of actions and procedures for healthier women is not adequate and needs to be well improved. However, these devices are not fool- proof. The summary of this paper is a partial result of a challenging problem faced during design and construction a fool-proof Smart Jacket for women’s safety using fabric sensors. The jacket consists of fabric Sensors, Accelerometer, Gyroscope and Magnetometer, which are strategically placed to record maximum variations in signal for minimum movement in subject’s body. The primary challenge is not in design or construction of a jacket, but in accurately classifying violent activity from animated activity. Both violent activity and animated activities have commonality in sensor excitation. There are subtle differences which needs to be extracted to train a Machine learning algorithm to learn these particular patterns. The process undertaken by other studies converges at a successful use of orientation sensors, fabric sensors and an effective system of integrating those sensors, in the form of a device, belt or a wearable gadget which can be worn by the subject and machine learning models have been applied on the data collected from these devices with varying degree of success. However, the major gap in the studies of the past is that none of the study led to a successful classification between normal rapid motion and violent assault. Thorough and granular study of previous research spanning all necessary spectrum of technology such as attack dynamics between an assaulter and the victim was undertaken to select the sensor position, leading to selection of motion parameters that would later become distinguishable features. The primary challenge in achieving distinct classification is the similarity in the data patterns of rapid motion of the body which are common in normal physical activities such as brisk action, playing games, running, or dancing. Using feature engineering and statistical analysis, new features have been created using a novel algorithm which uses the spatial-temporal parameters of the data and creates a sensor activation vector for a predefined length of time-stamped data. The training data has been collected from laboratory experiments in uncontrolled environment and the subjects were of different body types, thus ensuring that there would not be any bias in the data and the resulting model would perform well on the real data. The classical machine learning algorithms such as Decision Trees, KNN, SVM and Naïve Bayes have been used. The best result were obtained using Deep learning; the final result being a distinct classification between rapid movement from normal physical motion and rapid movement due to physical assault. |
