DEVELOPMENT OF COMPUTER VISION-BASED CONTAINER IDENTIFICATION AND STACKING SYSTEM ON GANTRY CRANE PROTOTYPE

• Main Author: Hilmi, Muhammad
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/66730
• Institution: Institut Teknologi Bandung

The heavy traffic of trade goods transportation via sea indicates the high demand for services in this sector of industry. With the increasing economic growth, the traffic density of this shipping service is also expected to increase in the future. Container loading and unloading activities at the port have a direct impact on this problem. Until now, these activities are carried out with a device called a gantry crane which is usually operated by a human. However, it will usually result in inconsistent operations which can reduce the productivity of loading and unloading activities and result in an increased dwell time of the containers at the port. The increase in dwell time will cause the traffic density and costs incurred at the port to be higher. Therefore, a gantry crane automation system is needed to minimize human involvement in the operation of gantry cranes for loading and unloading containers. In this final Project, a trolley position and crane swing angle control system and a computer vision-based identification system are implemented for the gantry crane prototype. Modeling of a 2-dimensional gantry crane prototype control system is implemented by using a PID controller. PID controllers are used to control the position of the trolley and PD controllers are used to control the swing angle of the crane. The tuning for the PID-PD controller is carried out in two stages, namely the weight of loss function and the gain parameter controller with the tuned weight applied. The tuning for the controllers is done with a PSO, FPA, SFS, and SA optimization. A computer vision-based identification system is done by static and dynamic cameras. The dynamic camera is used to perform serial number recognition with Tesseract OCR, object detection with SSDMobileNetV2 network, and a swing angle measurement. The static camera is used to perform depth estimation with various estimation models such as Monodepth2, FastDepth, MiDaS, and DPT. The integration of the control with the identification system is the key to various stacking maneuver executions. The variation of the stacking maneuvers that have been decided is lift-off, lift-on, and reshuffle maneuver. The results and analysis of the acquired data show that the tuning with PSO optimization results in the best simulation response performance of the trolley position and crane swing angle at 1 meter displacement with a cost of 0.0013. And the tuning with SA optimization has the best actual response on the gantry crane prototype with a cost of 0.0266. Serial number recognition is applied with image pre-processing which acquires a high accuracy of 97.35%. The trained object detection network has a good evaluation with mAP 0.8071 and AR 0.8487. And the DPT depth estimation model has the best accuracy to locate the container prototypes which reaches a value of 97.50%. Based on the test results in the execution of the stacking maneuver, the quality of the execution is based on the accuracy of the depth estimation that determines the setpoint of the control system and also a good system response for the movement of the trolley by the control system.