Development of vehicle speed estimation algorithm in video surveillance using deep learning | |
| Author | Keattisak Sangsuwan |
| Call Number | AIT Diss. no.ISE-24-04 |
| Subject(s) | Vehicles--Tracking Vehicles--Speed--Data processing Deep learning (Machine learning) |
| Note | A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Microelectronics and Embedded Systems |
| Publisher | Asian Institute of Technology |
| Abstract | Camera systems are widely used as a road traffic monitoring system, but the system does not have the ability to estimate the speed of the vehicles that are moving on the road. In order to use the system as a speed camera, a speed sensor such as RADAR is required to integrate with the system. However, thanks to the advanced development of computer vision technology, there is a potential possibility to integrate the speed estimation function into the camera system for vehicle speed estimation without using the speed sensor. In this work, a novel method to estimate speed of the vehicle in a traffic monitoring video without using the additional speed sensor is presented.The implementation of two-speed measurement models is proposed including the measurement of the traveling distance of the vehicle in a given unit of time and the measurement of the traveling time of the vehicle in a given unit of distance. Parameters of the models are received by defining four virtual intrusion lines on the road surface in the camera field of view. Then, YOLOv3, DeepSORT, GoodFeatureToTrack, and Pyramidal Lucas Kanade optical flow algorithms are implemented to detect and track the target vehicle. From the tracking data, pixel displacement between two consecutive frames (before and after the vehicle crosses the lines) is measured as the traveling distance. The number of frames that the vehicle uses while moving from the first line to the other lines is measured as the traveling time. These two parameters at each intrusion line are used as speed measurement metrics.The speed measurement metrics are solved by using tracking data of 20 vehicles at 4 different ground truth speeds measured by a laser speed gun. Then, the metrics are used to estimate the speed of 813 vehicles. The best accuracy is with Mean Absolute Error (MAE) of 3.38 km/h and Root Mean Squared Error (RMSE) of 4.69 km/h. The same dataset is tested on a Multilayer Perceptron Neural Network model. It can reach accuracy with MAE of 3.07 km/h and RMSE of 3.98 km/h. The open dataset BrnoCompSpeed is used to confirm our proposed method. The best accuracy from our method is with MAE of 1.81 km/h and this is agreed by its RMSE of 2.52 km/h. |
| Year | 2024 |
| Type | Dissertation |
| School | School of Engineering and Technology |
| Department | Department of Industrial Systems Engineering (DISE) |
| Academic Program/FoS | Microelectronics and Embedded Systems (MES) |
| Chairperson(s) | Mongkol Ekpanyapong; |
| Examination Committee(s) | Manukid Parnichkun;Dailey, Matthew N.; |
| Scholarship Donor(s) | Royal Thai Government;AIT Fellowship; |
| Degree | Thesis (Ph. D.) - Asian Institute of Technology, 2024 |