Collect SIFT features from multiple reference locations in the indoor environment. Use SIFT features to localize the agent. Experiments conducted on the Active Vision dataset show that SIFT features successfully localize the agent in most scenarios.

Find anchor-positive and anchor-negative distances using a pre-trained classification model. Use these distances as a measure of similarity among triplet images. Train an embedding network using an adaptive guided by the similarity. Experiments conducted on CARS196 dataset show that adaptive triplet margin loss performs better than its counterpart.

Extract importance or weights per location in the feature map using Bi-LSTM. Use score weighted feature for the final classification. Use max-score as a form of interpretation as to why the driver was classified as distracted. Train model end-to-end. Experiments conducted on State Farm Distracted Driver Detection dataset.