Dynamic scene smoke detection based on Box Plot and Fully Convolutional Network
Smoke has strong light transmission and blurry texture characteristics,and can be falsely detected as clouds and fog,leading to low recognition accuracy and substantial environmental interference of video-based single-stage smoke detection networks,posing challenges for actual field use.To address these issues,we develop a two-stage smoke detection algorithm based on Box Plot Background(BPB)and Full Convolution DNCNN(FCDN).In the first stage,we adopt a box plot statistical method to remove mobile interfering targets in the background queue and use maximum and minimum values in the background queue to establish a background model that is capable of adapting to dynamic scenes.In the second stage,we replace the fully connected layer with a convolutional layer to overcome input image size and shape limitations.The proposed two-stage smoke detection algorithm is tested on public datasets and self-collected data and demonstrated substantially reduced false detection rates.Using public camera data,the algorithm achieves a missed detection rate of 0.003 17,which is the lowest among the three methods in the experiment.The false detection rate achieves 58 and is comparable to the other two methods.The algorithm's Frames Per Second(FPS)outperforms the Gaussian mixture models and is slightly better than median filtering.In public smoke datasets and self-collected smoke datasets,the algorithm achieves the same missed detection rate as the other two methods,but significantly better false detection rates of 0.005 21 and 0.001 14,respectively.The algorithm's FPS remains the same as Gaussian mixture models(58)and median filtering(73).Together,the proposed two-stage smoke detection algorithm effectively minimizes the influence of environmental factors and significantly improves smoke detection accuracy.The experimental results demonstrate that the algorithm significantly reduces missed detection and false alarm rates.
万方数据
维普
