The global crime rate has been rising annually in recent years. Criminal activity has always been a major problem for society and the world. Rapid technological breakthroughs, especially in the areas of artificial intelligence and deep learning, have created new avenues for identifying criminal detection skills. Deep learning methods have shown great promise in evaluating a variety of data sources, including social media feeds, sensor data, and surveillance footage, to detect criminal activity, forecast occurrences, and support law enforcement in taking preventative action. This paper concentrated on using different CCTV footage from the Kaggle UCF crime dataset to identify criminal activity. The dataset includes 15 different sorts of crimes, with 30 movies in each category. Following that, the data was preprocessed. Next, we used CNN and YOLOv8 deep learning techniques to train the data. After that, the confusion matrix has been employed to calculate accuracy, precision, recall, f-measure, and MPF. The outcome demonstrates that we are putting forward a criminal detection system that uses CNN algorithms and variations to successfully identify criminals, even in crowded areas, with a good training accuracy and minimal training loss.