AMF-MOT: Multi-Object Tracking Based on Motion-Appearance Feature Fusion for Object Vehicle Loss and Occlusion

Accurately and continuously tracking object vehicles and obtaining their motion information are crucial for safe decision-making and planning in autonomous vehicles. However, the existing multi-object tracking algorithms suffer from degraded tracking accuracy and frequent object jumps when object vehicles are lost and occluded. To solve this problem, a vehicle multi-object tracking method (AMF-MOT) is proposed, which fuses the deep appearance and motion features of object vehicles. In the object selection module, the key objects influencing the ego vehicle's motion are selected by predicting the future dynamic security potential field. In the object tracking module, the vehicle re-recognition network and noise scale adaptive unscented Kalman filter algorithm are designed to extract deep appearance and motion features from sensor data before and after object loss and occlusion, respectively. Subsequently, the dynamic feature weight allocation algorithm is proposed to fuse the appearance and motion features, and based on this, the object similarity is calculated to associate vehicles to achieve continuous tracking. The experiment results on multiple benchmark datasets demonstrate that compared with the existing methods, the proposed AMF-MOT can not only achieve continuous tracking under object loss and occlusion but also improve the accuracy and stability of multi-object tracking.