As global populations increase, the analysis of crowds garners significant attention from both social and technological perspectives. The diverse nature of crowd behavior poses challenges in assessment, particularly in complex and unpredictable environments where manual monitoring falls short, posing risks to public safety and security. Various methods for detecting abnormal behaviors aim to address issues such as execution time, computational complexity, efficiency, robustness against occlusion, and generalizability. This study introduces an innovative approach to human activity-based anomaly detection and recognition in surveillance videos, leveraging deep learning (DL) techniques. Video input undergoes preprocessing for noise reduction and smoothing, followed by feature extraction using kernel local component analysis to monitor human activities. Subsequently, Bayesian network-based spatiotemporal neural networks classify the extracted features, revealing anomalies within the surveillance video dataset. Simulation results across different crowd datasets demonstrate performance metrics including mean average error, mean square error, training accuracy, validation accuracy, specificity, and F-measure. The proposed methodology achieves an MAE of 58%, MSE of 63%, training accuracy of 92%, validation accuracy of 96%, specificity of 89%, and F-measure of 68%.