Using Back-Propagation of Artificial Neural Network for Car Crash Identification

• Main Authors: Zih-Huan Shao, 邵子桓
• Other Authors: Chin-I Huang
• Format: Others
• Language: zh-TW
• Published: 2011
• Online Access: http://ndltd.ncl.edu.tw/handle/71731913954998919496

碩士 === 國立高雄第一科技大學 === 系統資訊與控制研究所 === 99 === After the car crash accident, it is often difficult to confirm who is main culprit or what happen. Base on this reason, this study propose different identifiable approach from the traditional statistical analysis or other methods for polices to catch more information of accident. In this study, employing Back-Propagation Neural Network based on a simple scene for car crash simulation with the PhysX physics engine. By this way, the computer simulation can calculate and catch the information of these vehicles and then infers the moment of accident. In our study, the accident problems divide into two cases: fixed point and non-fixed point. The first case employs the number of hidden layer is 110 by trial and error, the learning rate is 0.5, inertia factor is 0.5, and the error is 0.048 after iterating 25000 epoches. Then use 2016 original data to test the network model that accuracy is 95%. Furthermore, random test 168 data to test the model that accuracy is up to 85%. The second case employs the number of hidden layer is 340 by trial and error, the learning rate is 0.5, inertia factor is 0.5, the error is 0.0315 after iterating 10000 epoches. Then using 24192 original data to test the network model that accuracy is 97%. Furthermore, 2016 random test data to test the model that accuracy is up to 86%. In this research, we only discuss the fixed point as well as non-fixed point for car crash identification. The simulation results verify that BPN network model and its performance that can effectively identify the car crash indeed. In the future, increasing the complexity of vehicle crash problem such as on 3D non-uniform terrain and BPN network’s convergent property have to discuss.