TSF-Net:Abnormal Video Behavior Detection Framework Based on Two-Stream Appearance Motion Semantic Fusion
Video anomalous behavior detection aims to detect rare and random anomalous events,which usually deviate from expectations.Existing detection methods do not consider the appearance semantics and motion semantics of the events in conjunction,so they lead to poor detection results.Since the appearance semantics and motion semantics of normal events are highly consistent and correspond,while the semantic consistency of abnormal events is low.Based on this,this paper proposes a two-stream fusion hybrid network for abnormal behavior detection.The appearance semantics and motion semantics of events are recorded separately using a multi-scale fused self-encoder to generate corresponding reconstructed appearance frames and reconstructed motion frames.Afterwards,the two frames are fused to generate the final prediction frame.We introduce temporal attention to enhance the model's ability to represent motion semantics and capture normal behavior with effective regular motion.Experiments on three standard public datasets show that the method proposed in this paper can accurately detect anomalous events with a performance comparable to state-of-the-art methods.
video anomaly detectionTwo-Stream netprediction-based detection
NETL
NSTL
万方数据
