Human Abnormal Activity Pattern Analysis in Diverse Background Surveillance Videos Using SVM and ResNet50 Model

Today, almost all the places are observed by surveillance cameras. The aim is to monitor the activities in and around, especially for abnormal activities. But it requires manual assistance to watch/monitor. However, manual inspection is a monotonous job, that reflects on information lost. It shows the importance of an automatic abnormal activity detection system. This is considered a difficult task because of object overlapping, light variation, clutter background, camera angle, presence of various activity, and posture variations. It manipulates the human actions in videos. Hence, it is challgening to recognize abnormal human actions. This work examines the concert of the classical SVM and ResNet50 model among four datasets: The ‘SAIAZ’ (S tudents A ctivities I n A cademic Z one)-Corridor, ‘SAIAZ’-Open space, Classroom Violence from YouTube (cc), and Mixed Background Dataset (MBD). The ‘SAIAZ’ was created by student volunteers of our Institution, and MBD is a collection of selected frames from various videos. Abnormal actions like slapping, punching, kicking, running, and fighting is commonly extant in these datasets. Here, MBD is assumed to adopt various real-world situations. The SVM achieved the classification accuracy of 85%, 92%, 60%, and 44% on SAIAZ-Corridor, SAIAZ-Open space, Classroom Violence, MBD respectively. The ResNet50 achieved significant improvements in all the given datasets.