Real-Time Continuous Object Segmentation Using High Temporal Resolution Event Cameras via Random-Coupled Empowering Neural Network Optimized by Chef Leader Algorithm

The concept of event-based cameras has emerged as possibly helpful solution to the problem of high-temporal and low-latency capturing dynamic scenes and is therefore extremely applicable to the continuous object segmentation problem. Scarcity and asynchronous nature of event information, though, pose severe issues in feature description and precision of delineation. The given work proposes a systematic deep learning model on continuous object segmentation in event-based vision that is composed of the preprocessing, feature extraction, attention modeling, and optimization strategies. The framework uses EfficientNet-B3 to achieve the spatial features extraction, the attention mechanism to boost the temporal representation, the Random Coupled Empowering Neural Network (RCENN), to boost the interaction of features, and Chef Leader Optimization (CLO) to carry out the stable and effective optimization of the parameters. The proposed model is evaluated on EOS event-based data with regards to classification and segmentation. The experimental results indicate that the framework has excellent segmentation performance with the Intersection over Union (IoU) of 0.82, Dice coefficient of 0.86 and mean Intersection over Union (mIoU) of 0.81 and the classification performance is excellent. Ablation experiments and comparison are also effective in demonstrating the contribution of each aspect towards the enhancement of performance. Although the results are promising, the performance relies on the properties of datasets and experimental conditions and is yet to be demonstrated on more diverse real-life situations. Generally, the framework presented is a scalable and efficient solution to the event-based segmentation of dynamic environments.