Early identification of human suspicious activities, especially in academia, is very crucial for enhancing security and safety, while existing surveillance systems remain ineffective at contextual behavior analysis. This study proposes the HYDPL Algorithm that incorporates a novel pseudo-labeling approach and innovative image analysis to perform human suspicious activity detection. The HYDPL algorithm includes several stages of data processing and feature extraction that aim to enhance the detection of activities by increasing the model’s accuracy. To evaluate the effectiveness of the algorithm, two datasets were utilized: CampusWatch (Dataset-I), which consists of real-world scenarios specifically collected for this study from academia, targeting nine specific behaviors: Kicking, Punching, Running, Pushing, Smoking, Throwing, Jumping, Falling, and Talking, providing realistic representation of human activities, and HMDB51, named as Dataset-II, which is a collection of movies or scripted videos. The evaluation results of the model were outstanding for both datasets, as the model was successful in delivering 98% accuracy for Dataset-I and 97% for Dataset-II, highlighting the model’s ability to accurately detect real-world conditions. However, the study is limited to academia and only nine categories out of ten, with the tenth category being normal behavior. Future work will expand the dataset, explore advanced deep learning architectures, and implement real-time processing to enhance the model's applicability across various environments.