Using Baidu EasyDL to realize forest fire early warning and identification

Project Description
Business Background
According to statistics, in 2021, a total of 748,000 fires have been reported nationwide, with direct property losses as high as 6.75 billion yuan.Fire has become a frequent disaster that endangers people's life and property safety.For high-fire scenarios such as residences, gas stations, highways, and forests, if the fire cannot be detected and extinguished in time at the initial stage of the fire, it will lead to higher difficulty and loss of fire fighting.

Business Difficulty
Most fires occur in remote areas with large areas. Rapid feedback of early warning information is particularly important. Manual inspections alone are slow and easy to detect missed identifications.

Solutions
Through the EasyDL object detection task, a high-precision smoke and fire detection model can be trained in as little as 15 minutes. The AI ​​application can effectively prevent such problems and automatically detect the smoke in the monitoring area.and fire, help relevant personnel respond in time, and minimize casualties and property losses.

Data preparation
Data collection and import
It is recommended to collect data for real application scenarios, such as pictures of forest fires. If there are no corresponding rich samples, fires in the forest background can also be extracted from historical forest fire accident videosSmoke and tinder pictures.

Under normal circumstances, the use of drones for inspection is a common early warning method. The viewing angle of the picture during data collection must be consistent with the drone camera angle, as shown below:

The collected original images can be packaged and uploaded to the platform, and annotated using the built-in object detection and annotation tool on the platform.When collecting pictures, EasyDL intelligent data service provides data collection function, that is, after you download the data collection software to the local, by using AK, SK to link with the platform, the pictures collected on site can be sent back to the platform in real time for your next annotation.with training.

You can customize the camera data to be sent back according to the camera and time, and the picture will be captured and displayed according to the frame extraction rules you set.

In addition to capturing the local software through the camera and sending the data directly to the platform, you can also upload it locally or save the data to Baidu bos or online disk and upload it through a shared link.

Data annotation
Since a target detection model needs to be trained to detect the fireworks in the picture, the target detection template needs to be selected for labeling. When labeling, pay attention to all the smoke and fire in the pictures that need to be framed (Boxes can overlap), the detection box should contain the entire recognition object, and as far as possible not contain redundant background.

Tip: Due to the large number of interference samples for fireworks detection, it is very easy to cause false detection. There are many objects in life that are very close to fireworks, and it is difficult to distinguish (such as clouds, red lights, etc.), which is easy to cause false detection of the model.Therefore, it is recommended to also collect a certain amount of data as a negative sample.

Model training
Select the object detection task type, click to create a model, and customize the naming platform according to the actual business. The platform supports multiple deployment methods such as public cloud, local server, and small edge devices. For details, see: https://ai.baidu.com/ai-doc/EASYDL/dk38n33k4 In smoke and fire detection scenarios, we hope that the model can respond in time as soon as a fire occurs, and has high requirements for model inference speed, so it is recommended to useThe model is deployed to edge nodes closer to the monitoring site, which reduces network bandwidth and shortens service request latency.When choosing an algorithm, you can choose a high-performance algorithm to obtain faster inference speed.

Model deployment
Due to the high real-time requirements of the fireworks detection scenario, it is recommended to use the small edge device deployment method, which can be manually deployed by downloading the sdk offline or using the device-cloud collaboration function to directly deliver the deployment package to the edge device.

Performance optimization
It is a common business problem that you cannot get a good model performance after training once. You can continuously optimize your model performance in the following ways: By viewing the model evaluation report, you can find the distribution characteristics of identifying wrong or difficult cases,And targeted expansion of the data set of the corresponding scene, for example, in this case, the cases of fireworks identification errors are more concentrated in the foggy weather, due to the foggy weather, the fireworks cannot be accurately recognized or the fog is wrongly identified as smoke, you canCollect more data for this scene to expand the dataset, and return to the forecast service to collect more real scene data, and continue to iterate on the model effect.

FAQs
Question 1: How much data should I collect?

In terms of the number of data collection, firstly, it is necessary to ensure that the data volume of each label is not less than 50. In theory, the more labels, the better the model effect; the second is to try to ensure that the data volume of each label is not too different.That is to have a certain balance.

Question 2: Fire events do not occur frequently in real scenes, how to collect more training data?

In order to obtain rich and diverse violation scene data conveniently and quickly, open source datasets can be added (the open source datasets must match the real scene), or virtual images can be generated through data enhancement technology.

Question 3: After the model is deployed, fireworks need to be detected in the video stream, and the model can be called to obtain prediction results after video frame extraction. Is there a simple way with less development?

It is recommended to use the IEC intelligent edge console. Through IEC, the EasyDL SDK can be deployed locally, and the local and remote cameras can be visually connected. For details, see: https://ai.baidu.com/ai-doc/EASYDL/Gktuwc59w