Road Surface Object Recognition Based on Few Shot Learning

Pang, Wei Qin (2023) Road Surface Object Recognition Based on Few Shot Learning. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Full Text -Pang Wei Qin.pdf Restricted to Registered users only Download (1MB)

Abstract

Recognizing road surface objects is essential for the autonomous car to lower the daily accident rate. Road surface objects can be observed from a distance since they may have an impact on a vehicle's mechanics and create risky situations that endanger people's safety. Significant research has been done using both conventional and deep learning techniques to identify objects on the road surface. The data collection has collected a total of 2000 images for 5 classes of road surface objects by utilizing the smartphone’s camera and dashboard camera. In this study, a prototype network in few-shot learning is developed to address the issue of data scarcity by performing recognition on the little dataset. The results have been done on N-way K-shot classification by using different CNN models to evaluate the performance of the Prototypical network. The optimum point of the Prototypical network for N-way K-shot is determined using ResNet-18 model. Besides, the EfficientNet B2 and EfficientNet B1 models achieved highest accuracy of 83.88% and 88.63% in the 5-way 5-shot and 3-way 5-shot respectively. The performance of the Prototypical network was also compared to the Siamese network for k-shot and one-shot learning scenarios respectively on 3-way. The Prototypical network outperformed the Siamese network due to its ability to perform on different k-shot learning scenarios while the Siamese network is well-suited to perform on one-shot learning tasks.