AI DRIVEN SUSPICIOUS ACTIVITY RECOGNITION USING DEEP LEARNING

Abstract: The growing deployment of closed-circuit television (CCTV) cameras produces large volumes of video that are difficult to monitor continuously through human observation alone. Manual surveillance is vulnerable to fatigue, delayed response, and missed events, while basic motion detection cannot reliably distinguish ordinary movement from potentially suspicious activity. This paper presents a deep-learning-based framework for suspicious human activity recognition from surveillance videos using a hybrid Convolutional Neural Network (CNN) and Long Short- Term Memory (LSTM) architecture. Video frames are extracted and preprocessed using OpenCV. A pretrained CNN, such as MobileNet or InceptionV3, is intended to extract spatial features from individual frames, while an LSTM models temporal relationships across frame sequences. A classification layer predicts activity categories and can provide confidence scores to support alert generation. The framework is designed using Python and TensorFlow/Keras, with potential evaluation on public video datasets such as UCF-Crime or HMDB51. Since measured implementation results were not provided in the project report, this paper does not claim a specific accuracy, frame rate, or real-world performance. Experimental validation is required to establish effectiveness, false-alarm rates, and deployment suitability.

Keywords: Suspicious activity recognition, intelligent surveillance, deep learning, CNN-LSTM, computer vision, human activity recognition, video analytics