中央研究院 資訊科學研究所

In this talk, I will present our works in CVPR 2019 Nvidia AI City Challenge Workshop. My talk includes two topics, the first topic is Multi-Camera Vehicle Tracking. Due to the exponential growth of traffic camera net- works, the need for multi-camera tracking (MCT) for intelligent transportation has received more and more attention. The challenges of MCT include similar vehicle models, significant feature variation in different orientations, color variation of the same car due to lighting conditions, small object sizes and frequent occlusion, as well as the varied resolutions of videos. In this work, we pro- pose an MCT system, which combines single-camera tracking (SCT) and inter-camera tracking (ICT) which includes trajectory-based camera link model and deep feature re- identification. For SCT, we use a TrackletNet Tracker (TNT), which effectively generates the moving trajectories of all detected vehicles by exploiting temporal and appearance in- formation of multiple tracklets that are created by associating bounding boxes of detected vehicles. The tracklets are generated based on CNN feature matching and intersection- over-union (IOU) in every single-camera view. In terms of deep feature re-identification, we exploit the temporal attention model to extract the most discriminant feature of each trajectory. In addition, we propose the trajectory- based camera link models with order constraint to efficiently leverage the spatial and temporal information for ICT. The proposed method is evaluated on CVPR AI City Challenge2019 City Flow dataset, achieving IDF1 70.59%, which outperforms competing methods. By the way, our vehicle re-identification method is evaluated on CVPR AI City Challenge 2019 Track 2 dataset, achieving mAP of 79.17% with the second place ranking in the competition. The second topic is Traffic Anomaly Detection. Anomaly event detection on road traffic has been a challenging field mainly due to lack of training data and a wide variety of anomaly cases. In this paper, we propose a novel two-stage framework for anomaly event detection in road traffic based on anomaly candidate identification and starting time estimation of vehicles. First, we use Gaussian mixture models (GMMs) to generate the foreground mask and background image to identify the anomaly candidates. Foreground mask is used to produce the region of interest (ROI) to filter out the noise from the object detector, YOLOv3, in the background image. Then, we apply the TrackletNet Tracker (TNT) to extract the trajectory of anomaly candidate to estimate the anomaly starting time. Experimental results, with achieved S3 score performance of 93.62%, on the Track 3 testing set of CVPR AI City Challenge 2019 City Flow dataset, show the effectiveness of the proposed framework and its robustness in different real scenes.

Dr. Hung-Min Hsu graduated from the Department of Engineering Science & Ocean Engineering of the National Taiwan University. Dr. Hsu is serving in the Research Center for Information Technology Innovation. Now he is a visiting scholar in Department of Electrical and Computer Engineering, University of Washington. His research expertise in Research Center for Information Technology Innovation includes data mining and machine learning. He has participated in a number of projects at the Research Center for Information Technology Innovation, including the Taiwan E-learning and Digital Archives Programs (TELDAP) sponsored by the National Science Council of Taiwan (2009-2012), System Management and Content Retrieval Technologies for Supporting Cloud-based Digital Archive Systems and Services Project sponsored by Academia Sinica (2013) and Academia Sinica Digital Center: System Management and Content Retrieval Technologies for Supporting Cloud-based Digital Archive Systems and Services Project sponsored by the Academia Sinica, and the Ministry of Science and Technology, Taiwan(2014-2018). In University of Washington, his research topic is Mult-camera tracking. He is the Winner of Track 1 (City-Scale Multi-Camera Vehicle Tracking), Runner-up of Track2 (City-Scale Multi-Camera Vehicle Re-Identification) and Track 3 (Traffic Anomaly Detection) at AI City Challenge Workshop in CVPR’19.