Off-campus Michigan Tech users: To download campus access theses or dissertations, please use the following button to log in with your Michigan Tech ID and password: log in to proxy server

Non-Michigan Tech users: Please talk to your librarian about requesting this thesis or dissertation through interlibrary loan.

Date of Award


Document Type

Campus Access Master's Thesis

Degree Name

Master of Science in Electrical Engineering (MS)

Administrative Home Department

Department of Electrical and Computer Engineering

Advisor 1

Jeremy Bos

Committee Member 1

Michael Roggemann

Committee Member 2

Anthony Pinar


The pedestrian detection system is an important computer vision application and a part of Advanced Driver Assistance Systems (ADAS) which caters the need of automatically identifying pedestrians while driving a car. Since last few years, there have been significant advancements in the automobile industry to achieve a fully automated driving experience. As pedestrians are an important entity of the driving process, it is essential to maintain their safety while driving. Designing a pedestrian detection system involves classification and localization of a human shape, present in an input image or a video stream. However, detecting pedestrians in a moving environment is a hard task which has several challenges like different appearances, postures, different scales, various viewpoints, and frequent occlusions. Due to the existing neural network and handcrafted feature-based methods, to detect objects, these challenges can be handled. Although there exist a significant amount of developed pedestrian detection systems utilizing various methods, the comparison between handcrafted feature-based method and neural network-based method on the same testing platform and same testing data is yet to be performed.

This study aims to compare Histogram of Oriented Gradients (HOG) and Single Shot MultiBox Detector (SSD) methods on the same experimental setup. To achieve the goal, performance metrics such as detection rate, false positive rate, misprediction rate, the effect of varying detection window size on detection accuracy and speed of detection are taken into account. After testing the algorithms, it was identified that the size of the detection window and scaling factor plays an important role in maximizing the detection speed and accuracy of the Histogram of Oriented Gradients (HOG) detection system. Moreover, Single Shot MultiBox Detector (SSD) was better at detection accuracy than Histogram of Oriented Gradients (HOG). However, the time required to process Histogram of Oriented Gradients (HOG) was less as compared to Single Shot MultiBox Detector (SSD) algorithm when tested on a CPU. The performance of both the algorithms was subject to variation in illumination, size of the object and the processor used. However, these drawbacks can be reduced by using high computational power processors like GPU and FPGA. For future work, these algorithms can be tested on GPU and FPGA's with multiple object classes rather than only on pedestrians.