EZIKA, CHINWEIKE PRINCE (2026) DEVELOPMENT OF A HYBRID AUTONOMOUS CAR SIMULATION MODEL USING ARTIFICIAL NEURAL NETWORKS. Other thesis, Godfrey Okoye University, Enugu.
|
Text
Ezika Chinweike Prince.pdf Download (2MB) | Preview |
Abstract
Studies of autonomous vehicles have always taken place under controlled road conditions such as clearly marked lanes, consistent road traffic flow, and good maintenance of the road surface. Such studies cannot work under the conditions found in developing countries such as Nigeria, whereby roads are marked by unpredictable road agents, poor state of the roads, lack of lane discipline, and unpredictable human behavior. The research describes the creation of an autonomous car simulation model that works under these conditions. This architecture consists of an Expert System with deterministic Waypoint Path Planning as well as a Long Short-Term Memory recurrent neural network that acts as a reactive controller. The Unity simulation environment used for developing the front-end includes realistic vehicle dynamics and uses a sensor array with 12 distinct sensors, comprising proximity raycasts, object classifiers, kinematics sensors, offset from trajectory sensors, and a roughness detector. The LSTM network is trained via Supervised Imitation Learning from 91,000 expert demonstration examples created in simulation only, with mirroring applied to fix steering bias asymmetry issues. The inter-process communication between the Unity simulation environment and TensorFlow backend was achieved via Shared Memory (mmap), yielding latencies of under one millisecond, which represents a substantial improvement compared to the HTTP approach normally used. A thorough evaluation of the system through structured simulations involving open-road navigation, dense traffic conditions, rough road surfaces, and mixed road conditions showed that the system is capable of generating proper throttle, braking, and steering decisions, resulting in highly accurate steering with an MAE of 0.0198, or less than 0.69° of physical trajectory error per frame when compared to an expert driver. The system was able to generate decisions at 60 FPS+ with sub-millisecond latency even using general laptop-level consumer hardware without any specialized equipment. Through these findings, the current work proves that simulation-based hybrid expert-ANN control can be a useful and feasible avenue toward autonomous navigation for roads currently unexplored by global autonomous driving researchers.
| Item Type: | Thesis (Other) |
|---|---|
| Subjects: | Q Science > Q Science (General) |
| Divisions: | Faculty of Engineering, Science and Mathematics > School of Electronics and Computer Science |
| Depositing User: | MICHAEL MADUBUKO |
| Date Deposited: | 22 Jul 2026 12:25 |
| Last Modified: | 22 Jul 2026 12:25 |
| URI: | http://eprints.gouni.edu.ng/id/eprint/6011 |
Actions (login required)
 |
View Item |
