Download or read book Deep-learning-based Multiple Object Tracking in Traffic Surveillance Video written by Liqiang Ding and published by . This book was released on 2018 with total page pages. Available in PDF, EPUB and Kindle. Book excerpt: "Multiple object tracking (MOT) is an important topic in the computer vision. One of its important applications is in traffic surveillance for examining potential risks for traffic intersections and providing analysis of road usages. In this thesis, we propose a powerful and efficient model for solving MOT problems under traffic surveillance environments. The model solves MOT problems with the strategy of tracking-by-detection, and is flexible in tracking 11 categories of common road users from various altitudes and camera pitches. Moreover, it is an end-to-end solution that removes need of any further processes. There is no manual labeling required in the initialization step and objects can be tracked regardless of their motion states, which makes it possible to be applied in a large scale. We validate our model with multiple challenging datasets and compare its performance with other state-of-art methods. The evaluation shows our model can deliver satisfying results even though a simple data association algorithm is utilized. Optical flow and discrete Kalman filter achieve competitive performances in extracting and predicting motion states of objects. However, there are not many methods available to combine them with deep learning models to solve MOT problems. Our proposed model achieves object detection with a pre-trained deep learning detector, and then performs data association based on optical flow vectors, object categories, and object spatial locations. In order to improve the accuracy, a combination of techniques such as Gaussian mixture modeling is employed. To handle occlusion and lost track problems, a Kalman filter is introduced to extrapolate the motion and spatial states of an object in the next frame, so that the model can still keep tracking for a certain number of frames. Our model recovers a tracking trajectory by connecting the tracklet to a new detection response if it has similar optical flow and the same category in a region of interest. We comprehensively examine both detection and tracking stages with multiple datasets. Our detector delivers comparable results to several state-of-art methods, but with a faster processing speed. The tracking algorithm was evaluated in a benchmark test along with a few state-of-art methods. Compared to them, our model delivers competitive accuracy scores and usually achieves the best precision scores. In addition, we assemble a novel customized traffic surveillance dataset which contains videos taken under various weather, time, and camera conditions, and qualitatively test our model against it. The results demonstrate that our model well handles crowded scenarios or partial occlusions by generating smooth and complete tracking trajectories. Using a simple yet effective data association algorithm together with a Kalman filter, it serves as a powerful solution for MOT problems." --