Improved Spatial Invariance for Vehicle Platoon Application using New Pooling Method in Convolution Neural Network

Abstract

The imbalanced dataset is a prominent concern for automotive deep learning researchers. The proposed work provides a new mixed pooling strategy with enhanced performance for imbalanced vehicle dataset based on Convolution Neural Network (CNN). Pooling is crucial for improving spatial invariance, processing time, and overfitting in CNN architecture. Max and average pooling are often utilized in contemporary research articles. Both techniques of pooling have their own advantages and disadvantages. In this study, the advantages of both pooling algorithms are evaluated for the classification of three vehicles: car, bus, and truck for imbalanced datasets. For each epoch, the performance of max pooling, average pooling, and the new mixed pooling method was assessed using ROC, F1-score, and error rate. Comparing the performance of the max-pooling method to that of the average pooling method, it has been found that the max-pooling method is superior. The performance of the proposed mixed pooling approach is superior to that of the maximum pooling and average pooling methods. In terms of Receiver Operating Characteristics (ROC), the proposed mixed pooling technique is approximately 2 per cent better than the maximum pooling method and 8 per cent better than the mixed pooling method. Using a new pooling technique, the classification performance with an imbalanced dataset is improved, and also a novel mixed pooling method is proposed for the classification of vehicles.