This study explores the integration of smart city infrastructure to detect anomalous behaviors in crowds using state-of-the-art machine learning techniques. Crowds are defined as groups of people gathered for different purposes. These crowds may exhibit behaviors that can become chaotic, particularly in densely populated urban areas, which can have severe consequences for their safety. Hence, the ability to detect, identify, and predict these anomalies is a top priority for the police force and smart cities. Utilizing two distinct datasets, this study compares the effectiveness of using a Deep Learning model, specifically DenseNet, and an unsupervised Isolation Forest machine learning approach. The findings reveal that both models have their strengths and weaknesses depending on the size and complexity of the dataset. However, the Deep Learning model demonstrated superior accuracy in identifying complex crowd dynamics in both tested datasets, outperforming the Isolation Forest model in accuracy and other used metrics, achieving 97.91% accuracy on the smaller Motion Emotion dataset and 85.35% on the challenging UCF Crime, paving the way for developing more sophisticated crowd management tools for safer smart cities and keeping the safety of crowds' gathering in different situations.