Framework for Detecting Suspicious Vehicles Movements using Machine Learning Techniques

The increasing number of vehicles and changing safety risks are placing increasing demands on urban traffic monitoring. Real-Time detection of complex behaviors such as unpredictable driving, unauthorized stops, and route deviations is impossible with traditional systems based on human supervision or static rules. To detect suspicious vehicle movements, this research proposes a hybrid machine learning system that integrates deep learning, computer vision, and anomaly detection. Real-Time object detection is possible through YOLOv8, while Kalman filtering and DeepSort provide reliable multi-object tracking. Unsupervised anomaly detection uses CNN-LSTM models combined with autoencoders and DBSCAN clustering to detect unusual spatiotemporal patterns. On real-world traffic datasets, the system achieves 96.2% recognition accuracy and 93% F1 score, demonstrating minimal latency and high accuracy. It is ideal for applications in smart cities due to its scalability and edge compatibility. This method provides traffic police with real-time situational awareness and early accident detection by combining multimodal data with predictive analytics.