Object Detection Networks at the Edge: Hardware Optimization and Intelligent Transportation Systems Applications

Abstract

In recent years, deep learning (DL) and especially Convolutional Neural Networks (CNNs) have become a key component of many computer vision systems and applications due to their demonstrated capability to accurately process visual information. Object detection is one of the most important and challenging problems capable of being solved by DL; in general, higher object detection accuracy can be achieved by DL compared to other techniques. However, DL tends to require expensive GPUs or cloud-based services (the latter requiring a high internet bandwidth, latency, and other associated costs), making DL applications traditionally very expensive to implement in practice. This thesis emphasizes the optimization of DL computation for object detection and proposes designs of real-time Intelligent Transportation Systems (ITS) applications at the edge using hardware accelerators. To optimize DL computation, a novel stride 2 Winograd method is proposed for deep neural network (DNN) inference optimization. The proposed method provides new algorithms that trade expensive multiplications for cheap additions, thereby increasing efficiency by vastly decreasing computational complexity. The proposed algorithms support 1D, 2D, and 3D input for CNNs. Additionally, a novel Processing Element (PE) is proposed to process stride one and two convolution in the same FPGA module. These algorithms, implemented using a GPU and an FPGA, are demonstrated to provide better efficiency compared to regular convolution implementations for a variety of kernels. Additionally, three ITS applications are proposed. The first application is a License Plate Localization (LPL) system constructed using an architecture comprised of bottleneck depth-separable convolutions with inverted residuals. The second proposed application is a novel two stage real-time deep CNN recognition system for decals issued by the Commercial Vehicle Safety Alliance (CVSA). The third proposed application is a novel three stage real-time deep learning-based edge system for hazardous materials (HAZMATs) recognition. The designed custom object detection architectures for ITS applications are capable of highly accurate real-time prediction on edge computing devices (Intel, Google, and/or NVIDIA), thus providing enormous cost and performance advantages compared to current implementations.