Real Time Object Detection and Tracking Using SSD Mobilenetv2 on Jetbot GPU

The results of the experiments show that dataset size has a substantial impact on the accuracy of deep learning applications. For example, the Single Shot Detector (SSD) MobileNetV2 architecture, when trained on a dataset of 45,000 samples, achieved an impressive 97.8 % accuracy with an inference latency of 5.545 milliseconds (ms) on the Jetson Nano platform. In line with these discoveries, my research intends to combine the SSD MobileNetV2 model with Jetbot, a mobile robotics platform powered by NVIDIA Jetson Nano, to enable effective object tracking and detection. By using a pipeline that includes data collection, model training, inference, and motor control, the system optimizes performance by using pre-trained weights that may have been modified for certain datasets. Using tracking algorithms for constant movement monitoring, the system quickly recognizes and categorizes objects by utilizing real-time inference on the Jetson Nano. The system's practical usefulness in many circumstances, such as navigation, is shown via experimental validation, which also highlights the system's potential for enhancing autonomous systems and robotics.