复杂场景下基于特征融合的车辆跟踪

Vehicle Tracking Based on Feature Fusion in Complex Scenarios

针对恶劣天气(阴天、雨天、雪天)、空气污染(雾霾天)、黑夜等复杂的视频拍摄场景导致的目标与背景相似以及孪生网络没有模型更新的车辆跟踪问题,在确保算法实时性的前提下,以SiamFC(全卷积孪生网络)为基础对其特征提取、模型更新进行改进,提出了一种利用特征融合获取目标的表观特征和语义特征以及模型更新阈值优化的方法.试验结果表明:所提算法实现了在复杂场景下的车辆跟踪,特别是在处理目标与背景相似问题上,具有很强的泛化能力及鲁棒性.

Aiming at the vehicle tracking problem caused by complex video shooting scenes,such as bad weather(cloudy,rainy,snowy),air pollution(smoggy),dark night,etc.,that the target and background are similar and the twin network does not have model updates,under the premise of ensuring the real-time performance of the algorithm,the feature extraction and model update are improved on the basis of SiamFC(full convolutional twin network).A method of using feature fusion to obtain the apparent and semantic features of the target and the optimization of the model update threshold is proposed.The test results show that the proposed algorithm achieves vehicle tracking in complex scenes,especially in dealing with the similarity of the target and the background,with strong generalization ability and robustness.