Application of YOLO V5 for front-end loader detection in collision risk zones with reclaimers within a coal yard

Resumo

In port operations, large machines known as reclaimers are used. Some of these machines are capable of reclaiming stored products at a nominal capacity of 8,000 tons per hour. While utilizing these colossal machines enhances productivity and reduces costs, their large dimensions have an unintended side effect: an increased risk of material and personal damage in the event of a collision.
To mitigate this risk, a solution was proposed: training a YOLO V5 neural network to detect and alert the presence of other equipment within the defined risk zone. The target object for detection in this work is the front-end loader (pá carregadeira), chosen due to its high exposure to collision risk.
For training the neural network, 3,195 images of various types of front-end loaders were collected, depicting different scenarios and positions. Additional background images were included. Annotations for bounding boxes were created using a tool provided by the website app.roboflow.com. The dataset was divided into 80% training images and 20% validation images12.
For training the network, 3,195 images of various types of front-end loaders were collected, depicting different scenarios and positions. Additional background images were included. Annotations for bounding boxes were created using a tool provided by the website app.roboflow.com. Only a single class was defined, and the dataset was divided into 80% training images and 20% validation images.
The training process was conducted in a staggered manner to optimize GPU consumption on Google Colab. Initially, a test was performed with 50 epochs, a batch size of 30 samples, and input images in the 640x640 pixel format. Subsequently, based on the weights from the first training, a second test was conducted by increasing the number of epochs to 100. Following the same logic, a new test was performed with the following parameters: a batch size of 60 and tests at 100 and 200 epochs.
Remarkably, even in the first training, promising results were achieved: Recall: 0.94; mAP@0.5: 0.97; mAP@0.95: 0.77; Precision: 0.97
The second training yielded even better results when compared to the first:
Overall recall: 0.94; mAP@0.5: 0.97; mAP@0.5:0.95: 0.78414; Precision: 0.99
Upon analyzing some machine images, the confidence of the markings increased to approximately 90% certainty. In conclusion, the work thus far has demonstrated satisfactory and promising results, justifying the continuation of further studies and simulations