Vehicle intrusion detection in highway work zones using inertial sensors and lightweight deep learning

Highway work zones are prone to intrusion events that threaten workers' safety and disrupt operations. Existing intrusion detection systems often produce high false alarms, causing alarm fatigue and reduced responsiveness. To address this, a data-driven intrusion detection method is proposed to distinguish real vehicle intrusions from non-hazardous events using inertial measurement unit (IMU) sensors attached to traffic cones. Acceleration and angular velocity data were collected through field experiments involving vehicle collisions, manual handling, and wind displacement. After preprocessing and data augmentation, a lightweight Long Short-Term Memory (LSTM) model was trained and optimized for real-time performance on edge devices. Evaluation yielded a 96 % accuracy and a 97 % recall for actual intrusions. Resultant acceleration and angular velocity are recognized as key features. This cost-effective, scalable solution enhances safety by effectively identifying actual hazards, minimizing false alarms, and mitigating the negative impact of alarm fatigue in highway work zones.