Aiming at the hidden danger of fire caused by electric bicycles and gas tanks taken into elevators， an improved attention mechanism was proposed to detect dangerous goods in elevator scene， and a method based on the mechanism was proposed. With YOLOX-s as the baseline model， firstly， a depthwise separable convolution was introduced in the enhanced feature extraction network to replace the standard convolution， which improved the reasoning speed of the model. Secondly， an Efficient Convolutional Block Attention Module （ECBAM） based on mixed-domain was proposed and embedded into the backbone feature extraction network. In the channel attention part of ECBAM， two fully connected layers were replaced by a one-dimensional convolution， which not only reduced the complexity of Convolutional Block Attention Module （CBAM） but also improved the detection precision. Finally， a multi-frame collaboration algorithm was proposed to reduce the false alarms of dangerous goods’ intrusion into the elevator by combining the dangerous goods detection results of multiple images. Experimental results show that compared with YOLOX-s， the improved model can increase the mean Average Precision （mAP） by 1.05 percentage points， reduce the floating point computational cost by 34.1% and reduce the model size by 42.8%. The improved model reduces false alarms in practical applications and meets the precision and speed requirements of dangerous goods detection in elevator scene.