2nd ICAI 2022

International Conference on Automotive Industry 2022

Mladá Boleslav, Czech Republic

The main features of the system are listed below: 1. Correct classification of objects belonging to 3 classes: [‘milkrun’, ‘empty_ platform’, ‘plate_forklift’]. 2. Probability for class detection > 70%. 3. Object bounding box is secondary due to no practical use of precise bounding box in application – intersection over union >= 0.5. 4. Detection results saved in folder for verification. 5. Video processing with 2–5x real time speed. 6. Detection results allow to create a map of realized routes and deviations from the loop initiation frequency. 7. Use of 15 CCTV cameras as data source. 8. The sequence of detection of the ‘milkrun’ object on individual cameras defines the executed route. The routines analyzed by the software are shown in Figure 2. Three routines are defined: 1. 1<-> 4 ; WAREHOUSE <-> ASSEMBLY 2. 1-3-4 ; WAREHOUSE – FORMATION – ASSEMBLY 2. 2 <-> 4 ; BATTERY ELECTRODES WAREHOUSE – ASSEMBLY

Figure 2: Plant layout with routine points

Source: Own elaboration 2.2 Transfer Learning

Deep learning algorithms try to learn high-level features from large amounts of data, putting it ahead of typical machine learning. It uses a hierarchical feature extraction technique with an unsupervised or semi-supervised feature learning approach to extract data features automatically. Traditional machine learning methods, on the other hand, necessitate manually designing features, which places a significant burden on users. Deep learning can be defined as a machine learning representation learning algorithm

110