Traffic Sign Detection and Recognition Through fpga Approach and Deep Learning

Abstract

The objective of this Ph.D. work is Traffic Signs Detection and Recognition through FPGA Approach and Deep Learning . In developing countries like India, the road traffic flow rate has become a high level of concern as compared to developed countries. The autonomous vehicle on the road, or car drivers in normal vehicles should be able to detect obstacles, edges and should be able to see the traffic signs beforehand to avoid accident. The work in this thesis has been carried out to suggest solutions to these challenges by proposing and implementing the detection of traffic signs and their recognition. newline newlineIn this work, Simulink based traffic sign detection models for foggy, blurry images have been implemented for various traffic signs. The thesis further highlights the FPGA implementation of these models with a focus on area and power aspects. The efficient hardware implementation requires less area and less power consumption. These issues have been addressed by proposing the hardware efficient Simulink based traffic sign detection models implemented using Xilinx System Generator. To address the issues of resource utilization, Hardware-Software Co-Simulation models for detection of traffic signs have been created and have been implemented on Spartan-3E FPGA (a low cost hardware device). The on-chip power (including static and dynamic) consumption of all the Hardware-Software Co-Simulation models have been found using XPE tool and a detailed analysis has been carried out. newline newlineRecognition using deep learning technology can train the machine to do a particular task by relying on large databases. In this thesis, an object detection pre-trained model of Faster RCNN with Inception v2 as the detector has been trained and tested for traffic signs. To address the issues of image quality, a pre-trained deep learning model has been trained and tested on 30 and 50 speed signs captured in adverse conditions, with 10000 training steps and analysis in terms of loss factor has been carried out. The recognition rate has been calculate