With the development of the city scale, a more high-efficient traffic system is needed. Thus, the Intelligent Transport System (ITS) is proposed. This system aims to provide innovative services relating to different modes of transport and traffic management and enable users to be better informed and make safer, more coordinated, and 'smarter' use of transport networks. This thesis focuses on the detection based on Infrastructure camera sensors, especially the fisheye camera.

Two different methods are designed for this problem, the first one is a traditional method based on Background Subtraction. This method is verified at the intersection of University Ave and Chicago Ave, in Riverside, California, and in the CARLA simulator. It successfully detects vehicles the whole day regardless of the illumination weather and change. At the same time, a real-world vehicle dataset is collected for the second route. Also, in the CARLA experiment, our method significantly achieves improvement in terms of MOTA (multiple object tracking accuracy) and MOTP (multiple objects tracking precision). In the most complex scenario, our method outperforms the SOTA by 6.22\% on MOTP and 1.71 pixels on MOTA.

For the second route which is deep learning, there have been many efforts to apply deep learning to fisheye camera detection, but without a solid and large-scale fisheye image dataset, the neural network always has a bad detection performance. Thus, the second method tries to create a dataset that looks like the real world but is generated in a simulator (CARLA) to dismiss this problem. A real-world style dataset with ground truth labels is generated by modified AttentionGAN. Then deep learning object detection methods could be directly trained on the generated dataset. This project adopts YOLOV5 as the detection network. The final experiment result shows the trained network is able to detect all vehicles in verify part of the generated dataset and part of vehicles in the real-world dataset.