With the rapid growth of urban populations and increasing urbanization in major metropolitan areas, public security challenges have intensified, necessitating innovative solutions for effective crime prevention. Traditional statistical models often struggle to capture the complex spatio-temporal patterns of criminal activity, limiting their predictive capabilities. In contrast, proactive strategies such as predictive policing have shown significant potential in reducing crime rates and optimizing resource allocation in urban environments. Convolutional Neural Networks (CNNs), widely recognized for their effectiveness in pattern recognition tasks, offer a novel approach to modeling crime dynamics through heatmap analysis. This paper proposes a CNN-based framework for predicting high-risk areas using historical crime data transformed into multidimensional tensors. The methodology involves pre-processing georeferenced crime records into a spatio-temporal heatmap, training a CNN model with weighted loss functions, and evaluating its predictions against real-world data. Experimental results demonstrate the model’s ability to identify crime hot spots while reducing false positives. However, validation tests highlight computational and generalization challenges, underscoring the need for further refinements. These findings suggest that CNNs can serve as a valuable tool for enhancing urban security planning, while emphasizing the importance of hybrid approaches to improve the spatial accuracy and reliability of crime prediction.