Optimization Of Neural Network Architectures for Real-Time Object Detection in Autonomous Vehicles Using Variational Mimetic Operator Networks
K Kodeeswari, P. Loganathan, P. Narmatha, Rakhi Mutha, F. V. Jayasudha, V. Savitha · 2024
The Research focus lies in the optimization of neural network structures specifically for real-time object recognition applications in autonomous vehicles, which is accomplished through the use of Vibrational Mimetic Operator Networks (VMON) as the technology baseline. By combining variation principles with mimetic approaches, VMON achieves improved generalization and computational efficiency in the solution process. Due to tight latency requirements in autonomous vehicle applications, conventional neural architectures frequently find it challenging to strike a balance between speed and detection performance. To this end, harness dynamically adaptable layers and parameter pruning within framework, enabling efficient inference without sacrificing accuracy. Intersperse the architecture with both localized and global feedback mechanisms that dynamically renormalize weights in real time, minimizing response lag of system even more. Large-scale experiments show that VMON beats baseline detectors, such as YOLO and Faster R-CNN on latency, with up to 35% reduction in processing time in average autonomous navigation scenarios. Taken together, results highlight the ability of VMON to set new benchmarks in efficiency for object detection and pave the way for safer and more reliable autonomous vehicle systems. This will be followed by research on the scalability of this methodology to various vehicular platforms and conditions of real-world traffic.