Aim Fires are a serious threat to people’s lives and property. Detecting fires quickly and effectively and extinguishing them in the nascent stage is an effective way to reduce fire hazards. Currently, deep learning-based fire detection algorithms are usually deployed on the PC side. Methods After migrating to small embedded devices, the accuracy and speed of recognition are degraded due to the lack of computing power. In this paper, we propose a real-time fire detection algorithm based on MobileNetV3-large and yolov4, replacing CSP Darknet53 in yolov4 with MobileNetV3-large to achieve the initial extraction of flame and smoke features while greatly reducing the computational effort of the network structure. A path connecting PANet was explored on Gbneck(104, 104, 24), while SPP was embedded in the path from MobileNetV3 to PANet to improve the feature extraction capability for small targets; the PANet in yolo4 was improved by combining the BiFPN path fusion method, and the improved PANet further improved the feature extraction capability; the Vision Transformer model is added to the backbone feature extraction network and PANet of the YOLOv4 model to give full play to the model’s multi-headed attention mechanism for pre-processing image features; adding ECA Net to the head network of yolo4 improves the overall recognition performance of the network. Result These algorithms run well on PC and reach 95.14% recognition accuracy on the public dataset BoWFire. Finally, these algorithms were migrated to the Jeston Xavier NX platform, and the entire network was quantized and accelerated with the TensorRT algorithm. With the image propagation function of the fire robot, the overall recognition frame rate can reach about 26.13 with high real-time performance while maintaining a high recognition accuracy. Conclusion Several comparative experiments have also validated the effectiveness of this paper’s improvements to the YOLOv4 algorithm and the superiority of these structures. With the effective integration of these components, the algorithm shows high accuracy and real-time performance.

Falls cause great harm to people, and the current, more mature fall detection algorithms cannot be well-migrated to the embedded platform because of the huge amount of calculation. Hence, they do not have a good application. A lightweight fall detection algorithm based on the AlphaPose optimization model and ST-GCN was proposed. Firstly, based on YOLOv4, the structure of GhostNet is used to replace the DSPDarknet53 backbone network of the YOLOv4 network structure, the path convergence network is converted into BiFPN (bidirectional feature pyramid network), and DSC (deep separable convolution) is used to replace the standard volume of spatial pyramid pool, BiFPN, and YOLO head network product. Then, the TensorRt acceleration engine is used to accelerate the improved and optimized YOLO algorithm. In addition, a new type of Mosaic data enhancement algorithm is used to enhance the pedestrian detection algorithm, improving the effect of training. Secondly, use the TensorRt acceleration engine to optimize attitude estimation AlphaPose model, speeding up the inference speed of the attitude joint points. Finally, the spatiotemporal graph convolution (ST-GCN) is applied to detect and recognize actions such as falls, which meets the effective fall in different scenarios. The experimental results show that, on the embedded platform Jeston nano, when the image resolution is 416 × 416, the detection frame rate of this method is stable at about 8.33. At the same time, the accuracy of the algorithm in this paper on the UR dataset and the Le2i dataset has reached 97.28% and 96.86%, respectively. The proposed method has good real-time performance and reliable accuracy. It can be applied in the embedded platform to detect the fall state of people in real time.

No abstract

Forest fires often have a devastating effect on the planet’s ecology. Accurate and rapid monitoring of forest fires has therefore become a major focus of current research. Considering that manual monitoring is often inefficient, UAV-based remote sensing fire monitoring algorithms based on deep learning are widely studied and used. In UAV monitoring, the size of the flames is very small and potentially heavily obscured by trees, so the algorithm is limited in the amount of valid information it can extract. If we were to increase the ability of the algorithm to extract valid information simply by increasing the complexity of the algorithm, then the algorithm would run much slower, ultimately reducing the value of the algorithm to the application. To achieve a breakthrough in both algorithm speed and accuracy, this manuscript proposes a two-stage recognition method that combines the novel YOLO algorithm (FireYOLO) with Real-ESRGAN. Firstly, as regards the structure of the FireYOLO algorithm, “the backbone part adopts GhostNet and introduces a dynamic convolutional structure, which im-proves the information extraction capability of the morphologically variable flame while greatly reducing the computational effort; the neck part introduces a novel cross-layer connected, two-branch Feature Pyramid Networks (FPN) structure, which greatly improves the information extraction capability of small targets and reduces the loss in the information transmission process; the head embeds the attention-guided module (ESNet) proposed in this paper, which enhances the attention capability of small targets”. Secondly, the flame region recognized by FireYOLO is input into Real-ESRGAN after a series of cropping and stitching operations to enhance the clarity, and then the enhanced image is recognized for the second time with FireYOLO, and, finally, the recognition result is overwritten back into the original image. Our experiments show that the algorithms in this paper run very well on both PC-based and embedded devices, adapting very well to situations where they are obscured by trees as well as changes in lighting. The overall recognition speed of Jeston Xavier NX is about 20.67 FPS (latency-free real-time inference), which is 21.09% higher than the AP of YOLOv5x, and are one of the best performance fire detection algorithm with excellent application prospects.

Blueberries are grown worldwide because of their high nutritional value; however, manual picking is difficult, and expert pickers are scarce. To meet the real needs of the market, picking robots that can identify the ripeness of blueberries are increasingly being used to replace manual operators. However, they struggle to accurately identify the ripeness of blueberries because of the heavy shading between the fruits and the small size of the fruit. This makes it difficult to obtain sufficient information on characteristics; and the disturbances caused by environmental changes remain unsolved. Additionally, the picking robot has limited computational power for running complex algorithms. To address these issues, we propose a new YOLO-based algorithm to detect the ripeness of blueberry fruits. The algorithm improves the structure of YOLOv5x. We replaced the fully connected layer with a one-dimensional convolution and also replaced the high-latitude convolution with a null convolution based on the structure of CBAM, and finally obtained a lightweight CBAM structure with efficient attention-guiding capability (Little-CBAM), which we embedded into MobileNetv3 while replacing the original backbone structure with the improved MobileNetv3. We expanded the original three-layer neck path by one to create a larger-scale detection layer leading from the backbone network. We added a multi-scale fusion module to the channel attention mechanism to build a multi-method feature extractor (MSSENet) and then embedded the designed channel attention module into the head network, which can significantly enhance the feature representation capability of the small target detection network and the anti-interference capability of the algorithm. Considering that these improvements will significantly extend the training time of the algorithm, we used EIOU_Loss instead of CIOU_Loss, whereas the k-means++ algorithm was used to cluster the detection frames such that the generated predefined anchor frames are better adapted to the scale of the blueberries. The algorithm in this study achieved a final mAP of 78.3% on the PC terminal, which was 9% higher than that of YOLOv5x, and the FPS was 2.1 times higher than that of YOLOv5x. By translating the algorithm into a picking robot, the algorithm in this study ran at 47 FPS and achieved real-time detection well beyond that achieved manually.