The real-time detection of fire protection facilities hidden trouble is of great significance to the prevention of fire. A real-time detection method for fire facility hidden trouble based on improved YOLOv3 algorithm is proposed. The YOLOv3 basic convolutional neural network model is deleted properly. The non-overlapping pooling layer is changed into overlapping pooling layer to avoid overfitting. The BN (Batch Normalization) layer is added between the convolution layers to speed up the convergence rate and improve the stability of the network. Each standard convolutional layer is split into a bottleneck layer, which increases feature extraction and significantly reduces network parameters. Replace the ReLU activation function with the Swish activation function to avoid the destruction of the feature. At the same time, the number and size of anchors are determined by K-mean++ algorithm, which makes the object positioning more accurate and the detection accuracy is higher. In view of the possible various fire protection facilities hidden trouble, under the different illumination, through the robot movement, carries on the real-time detection to the potential fire protection facilities. The IoU (Intersection-over-Union) of the improved algorithm is 85.23%, the mAP (mean average precision) is 96.89%, the FPS (Frame Per Second) is 30.47, which is 2.56%, 6.68% and 6.72 higher than that of the YOLO V3 algorithm respectively. And physical experiments in different indoor environments, the results show that the improved algorithm has low requirements for computer hardware, high detection accuracy and good real-time performance.
展开▼
