The NovelHAP Algorithm for Human Crime Activity Prediction in Video Frames using Kalman Filter and Template Matching

Human Activity Prediction (HAP) interprets human motion and future activity. Surveillance systems must use HAP techniques to deter crimes and dangerous behavior. The primary objective of this study is to forecast human criminal acts in video frames, encompassing shooting, fighting, and kicking. This study introduces the NovelHAP algorithm, built upon the Kalman Filter Algorithm (KFA), MediaPipe (MPP) framework, and Template-Matching Statistical Approaches (TMSA). KFA is used to predict a video frame that contains a very near state (or position) of one human Bounding Box (BB) and its centroid position compared to another in the video frames. When the two humans are very close to each other, the activity of the human can be predicted. The predicted frames have been considered as input to the MPP Framework to find the landmarks and posture of the human. This study uses the values of key landmarks Left_Wrist, Right_Wrist, Left_Hip, Right_Hip, Left_Ankle, and Right_Ankle of human postures in video frames. The Euclidean distances from D1 to D10 among the landmark vectors of video frame postures (training vectors) and template vectors (test vectors) are used as the input for TMSA. TMSA measures the similarity between training and test vectors to predict and classify human crime activities. The empirical analysis revealed that the accuracies for HAP on the UCF-101 dataset are 95% for small data samples, and 96.01% for large data samples, respectively.