Batch-mode active learning for traffic sign recognition

Cognitive vehicles face a huge variety of different objects when perceiving their environment using video cameras. Their capability of recognizing certain types of objects in a robust way is tightly coupled to successfully discriminating positive samples from negative ones (background objects). Machine learning methods trained on huge data sets can accomplish this, but to reduce labeling costs and classification runtime it is desirable to minimize the number of training samples needed. We examine the feasibility of active learning paradigms to achieve this goal for traffic sign recognition and propose a batch-mode multi-class active learning query strategy for support vector machines: Angular diversity ranking is a weighted combination of relevance of unlabeled samples with a diversity measure. A quantitative evaluation on a large data set with more than 30,000 samples shows clear benefits of our query strategy for traffic sign recognition, track-pruned class-biased angular diversity ranking, compared to uncertainty sampling based active learning as well as passive learning.