Impact of quality of training dataset on pedestrian detection for autonomous driving
In autonomous driving, the perception task is a key piece requiring object detectors like pedestrian detection and vehicle detection among other. These object detectors are typically implemented using machine learning algorithms involving pipeline of classifiers or using deep learning networks. Special considerations are required for training dataset for each classifier in the pipeline. The performance of generated classifier models depends not just the size of the dataset but also the on quality of the training dataset in terms of coverage of scenarios. This work also presents the impact of ratio of positive and negative training samples sizes towards improving the performance of different types of classifiers in the pipeline.
Sudipta Bhattacharjee is Associate Solution Architect at the KPIT Technologies Ltd. for ADAS practice. He did his Ph. D. from Indian Institute of Technology, Kharagpur, India. He has more than 10 years of experience for conducting industrial R&D work. His areas of work includes computer vision techniques, machine learning, data analytics, wireless sensor network, electronic system design & integration. He has filed 5 patents, and published 1 book, 9 journals and 11 conference papers. He is the winner of scholarship A for 2014 ISRM International Symposium (ARMS8), Sapporo, Japan. He is the peer reviewer of IEEE transactions on industrial electronics, Journal of optics & lasers in engineering, IEEE sensor journal and IEEE wireless communication magazine.
ACM Chapters Computer Science in Cars Symposium 2017