Evaluation of SSD Architecture for Small Size Object Detection: A Case Study on UAV Oil Pipeline Monitoring

Abstract

Oil pipeline monitoring using Unmanned Airborne Vehicles (UAV) can be done by utilizing Deep Learning. Deep Learning can be used to automatically detect harmed or unauthorized objects near the pipeline for further action by the authority. Input video in the pipeline area taken from the UAV has unique characteristics. It has low resolution with dense composition object in the image. The detected object also has a small scale as the objects are far away from the UAV. Thus, the selection of the Deep Learning algorithm is important to get a desirable result with the following conditions. Single Shot Multi-Box (SSD) is one of the popular Deep Learning algorithms with fast calculation compared to others and suitable for real-time object detection. Previous works on this topic using low to medium altitude dataset (20–200 m). This paper provides an evaluation of SSD implementation to detect vehicles on high-altitude dataset (300 m). As much as 2482 dataset is fed into SSD architecture and trained to detect 3 class of vehicles. The result shows the mAP and mAR are 0.026360 and 0.067377, respectively. However, the low lost function value shows that the model is able to classify the object correctly. In conclusion, the SSD cannot process low density information to correctly locate the object.