SSD Mobilenetv1-Based Pedestrian Detection System in Limited Environmetn Using Normalized 360° Images

Abstract

Along with the times, every car manufacturer always creates its newest products that are
more sophisticated. This idea then gave birth to the concept of an autonomous electric vehicle (KLO .) This purpose is intended to always present vehicles that can meet consumers' growing tastes while also being environmentally friendly. The presence of autonomous electric vehicles will certainly be experienced by Indonesia, whose people have started to rely on car transportation. Therefore, this situation requires us to be prepared to face the era of Mobility in Society 5.0, where we must master the supporting technology. Autonomous electric vehicles can be realized if the system can detect objects properly. Therefore, a deep learning-based pedestrian detection system was developed and utilized 360° images in this study.
The object detection software system was built using the Single Shot Multibox Detector (SSD)
MobilenetV1, while the hardware used for this development is Jetson AGX Xavier. The development process started from taking a normalized 360° image containing pedestrian information in the Nurtanio University campus area which was then used as a dataset and test data, training the MobileNetV1 SSD with the dataset (19038), and testing the trained software model in real-time and offline. Offline test results on 735 360° images in daytime conditions show that 55.5% of images can be detected ideally, while of 595 360° images in afternoon conditions, 51.2% of images can be ideally detected. In real-time testing, 98% of pedestrians are inevitably seen during the day, while only 95% in the afternoon. The average processing time on an image in daytime conditions is 32.81283 ms if using the CPU, while if using the GPU, it is 32.79766 ms. For an image with the same information in the afternoon conditions, the processing time is 37.42598 ms if using the CPU, while if using the GPU, it is 37.45174 ms.