This project plans to develop a vehicle detection and tracking model to detect small vehicles from low-quality high-altitude roadside images. It can effectively estimate vehicle flow and related statistics which can be then used for traffic incident data simulation and generation, even though low-quality video data flows are used.  This project will use deep learning to develop a new lightweight architecture, that more accurately detects small objects from low-quality traffic video data. Then, traffic flow at each interaction can be well estimated and used for traffic conflict simulation and hot spot identification.  The simulation result of vehicle flow can provide important information for reference in subsequent traffic management and safety analysis.  Thus, effectively evacuating the intersection traffic quickly can be achieved.

This technology develops a light-weight deep learning network architecture, that uses aerial images to analyze traffic flow. Due to the high altitude, the objects on the ground are often less than 20 * 20 pixels. In this case, even the famous method of the object detector YOLO v3 or SSD none of them can be effectively detected. This technology will develop a new light-weight network architecture and try to solve some problems with pooling. Not only the speed is improved, but also the accuracy of the detection of small objects is improved.

SSAM can be used as a basic set of developing diagnostic tools for high accident traffic risk location. Diagnose traffic conflicts through basic traffic information at intersections, thereby assisting traffic management units and road designers as preventive traffic safety improvements.  In addition, with the lightweight network architecture, we can develop a RL-based traffic sign control system  to deal with the problem of traffic jam.