Sudden change detection on roads using multiple image processing on autoencoder | |
| Author | Pongsaton Mondee |
| Call Number | AIT Thesis no.ISE-21-18 |
| Subject(s) | Deep learning (Machine learning) Optical radar Image processing--Digital techiniques |
| Note | A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Mechatronics |
| Publisher | Asian Institute of Technology |
| Abstract | Nowadays, most road accidents are caused by drivers more than the vehicles and environmental conditions combined. One solution is to reduce the number of human deci sions in driving by using the autopilot system or driving assistant to help driving more secure. In order to make the decision, autonomous driving vehicles need to be aware of the surrounding moment for calculating and determine to drive precisely and safely. This thesis focuses on computer vision fields with the ability to interpret images of the environment around the car. Instead of using a LiDAR scanner, this using the front car’s camera and autoencoder neuron network technologies combined can also be used to identify any anomaly moment that occurs. In this study, the score is used to indicate anomaly events in the footage. An anomaly score is created by the fact that an abnormal moment occurs infrequently when the point of the system is to learn to reconstruct the image sequences by using an autoencoder network implement with Conv-LSTMs. The anomaly moment will get a greater error from reconstructing sequences. The system is also equipped with 2 subs networks for transform raw images into semantic segmentation images and dense optical flow im ages. This preprocess is for reducing the complexity of the images before sending them to the Conv-LSTMs autoencoder. To make a system that can interpret anomaly moments on the road, many parameters are trained and tested by the system such as input image size, sequences, number of color channels, and other parameters. The system is also unsupervised trained between images from the city and the rural scenarios. From the test results, the best system of road anomaly detection is quite well detecting the anomaly events on the road in the three different tests set, city dataset, rural dataset, and Extest dataset with the best F score of 64.12%, 63.53%, and 66.67 % correspondingly and the overall score is 62.12 %. The best system is made by using HD resolution as input images in the preprocess network and then reducing the sequenced image’s dimension down to 256x256 pixels with 1 color channel in segmentation while has 3 color channels in optical flow for the Conv-LSTMs network. In the calculation of the anomaly event process, the anomaly is computed using the mahalanobis distance on a reconstruction score from the 2 different models, dense optical flow, and segmentation with thousand steps and thousand initial points. The network uses around 7-10 minutes in order to process a 1-minute-long video and can detect sudden changes around every 15-20 seconds. |
| Year | 2021 |
| Type | Thesis |
| School | School of Engineering and Technology (SET) |
| Department | Department of Industrial Systems Engineering (DISE) |
| Academic Program/FoS | Industrial Systems Engineering (ISE) |
| Chairperson(s) | Mongkol Ekpanyapong |
| Examination Committee(s) | Manukid Parnichkun;Dailey, Matthew N. |
| Scholarship Donor(s) | Royal Thai Government Fellowship |
| Degree | Thesis (M. Eng.) - Asian Institute of Technology, 2021 |