Innovative convolutional neural networks for object detection in autonomous driving systems

In this thesis, first, the results of the following novel studies are presented that reveal the limitations of using Convolutional Neural Networks (CNN) in object detection modules of Autonomous Driving Systems (ADSs): (i) a comprehensive investigation of the state-of-the-art of CNNs for image classification, object detection, and ADSs to de-termine the real-time constraints in a synergetic way; (ii) a trade-off analysis from a human vs machine and accuracy vs complexity perspective for novel CNN designs; and (iii) a novel analysis of commonly used convolution and operation types in CNNs as well as including a timing analysis aimed at assessing their impact on CNN performance. Based on the outcome of these studies, then a novel design approach is proposed, which fuses three strategies to build a new Sparse-Split-Parallelism (SSP) design framework. The SSP framework can be applied to existing block-based CNN models to lighten their computing budget, while maintaining the same performance. As the first strategy, a new type of sparse skip connections is introduced as a block-level strategy. At the module-level, a proportional channel split operation is employed, which characterizes the accu-racy and model size trade-off. As the third layer-level strategy, an equal number of input and output channels in layers is selected so that the degree of parallelization is increased for faster inference.

To verify the efficacy of the proposed SSP framework, a comprehensive evalua-tion is conducted using three highly competitive visual recognition tasks employing ImageNet, MS COCO, and KITTI test datasets, including a real-time experiment as well. It is shown that as a result of the application of the SSP framework, the SSP-enabled CNNs for state-of-the-art CNNs, SSP-DenseNet, SSP-ResNet, and SSP-ShiftNet, outper-form their ancestors featuring ~3x lower number of parameters, ~2x lower floating-point operations (FLOPs), and ~2x faster inference performance, while attaining the same level of accuracies.