This study explores advanced deep learning techniques for detecting suspicious activities in video surveillance systems. A thorough review of existing literature highlights the importance of intelligent video analytics in enhancing security measures and public safety. The research emphasizes the effectiveness of deep convolutional neural networks combined with recurrent architectures for feature extraction and temporal analysis of video data. To address current limitations, a novel framework is proposed, integrating an object detection model with custom logic for real-time monitoring of crowd density, headcount tracking, and detection of anomalous activities. The solution includes preprocessing strategies, precise training methods, and instant SMS alerts for overcrowding or the presence of prohibited items. Additionally, the study outlines a systematic approach for acquiring and annotating datasets that capture suspicious behaviors in various environments, aiming to improve model generalization. The proposed system aims to provide authorities with actionable insights for rapid response, thereby preventing security breaches and ensuring responsible oversight of public spaces. The effectiveness of the approach is demonstrated through benchmarking analysis on key performance indicators such as accuracy, latency, and reliability, advancing the field of automated suspicious activity recognition in video surveillance.