Analysis and Comparison of FPGA-Based Histogram of Oriented Gradients Implementations

Abstract

One of the commonly-used feature extraction algorithms in computer vision is the histogram of oriented gradients. Extracting the features from an image using this algorithm requires a large amount of computations. One way to boost the speed is to implement this algorithm on field programmable gate arrays, to benefit from flexible designs such as parallel computing. In this paper, we first, provide a summary of the steps of the histogram of oriented gradients algorithm. We then survey the implementation techniques of the histogram of oriented gradients on field-programmable gate arrays in the past decade. We group the different techniques into four main categories and analyze various enhancement methods in each category. The first group is the optimization of the algorithm computation which involves the steps of input selection, magnitude calculation, orientation and bin assignment, and normalization. The second category is data manipulation techniques which include numerical representation, data flow modification, and memory optimization. The third group contains modified features based on the histogram of oriented gradients and their hardware implementation, and the fourth one is the implementations in hardware-software co-design of the algorithm. We compare the different implementations using a speed metric called pixels per clock cycle, and resource utilization. Finally, we provide design summary tables for efficient implementation with respect to the speed metric, accuracy, and resource utilization. INDEX TERMS Histogram of oriented gradients, field programmable gate arrays, hardware acceleration, hardware design. KIN FUN LI is the Director of the two highly sought-after professional master of engineering programs in telecommunications and information security (MTIS) and applied data science (MADS) at the University of Victoria, Canada, where he teaches both hardware and software courses in the Department of Electrical and Computer Engineering. He dedicates his time to instructing and researching in computer architecture, hardware accelerators, education analytics, and data mining applications. He is actively involved in the organization of many international conferences, including the biennial IEEE Pacific Rim in Victoria and the internationally held IEEE AINA. He is also a passionate supporter and participant in numerous international activities to promote the engineering profession, education, and diversity.

A novel feature based algorithm for soil type classification

“…Srisamosorn et al [50] developed an improved HOG named Fisheye HOG (FEHOG) where they reduced the computational time in several stages. Ghaffari et al [19] performed a comparative study for FPGA based HOG feature implementations and they mentioned that FPGA-HOG features are computationally faster than the traditional HOG. Sharma et al [47] proposed a transformed version of HOG where they used only magnitude information rather than orientation information, and they also claimed it to be faster than the traditional HOG.…”

A Novel Hardware–Software Co-Design and Implementation of the HOG Algorithm

“…We have surveyed different methods for hardware implementation of the HOG algorithm, including an extensive review of methods with hardware–software co-design in our previous work [ 16 , 17 ]. In this section, first we briefly review the recent work implementing the HOG algorithm fully on hardware.…”

“…As mentioned in [ 17 ], the original HOG algorithm requires 2 × W × H multiplication operations (for computing the square of the gradients twice for each pixel), W × H additions (once for each pixel), and W × H square root operations (once for each pixel) for computing the magnitude of gradients, where W is the width and H is the height of the image window. In our implementation, we simplified the magnitude computation by just performing W × H additions (for adding the absolute values once for each pixel) and 2 × W × H inversion operations (for absolute value of the gradients twice per pixel).…”

“…It is important to note that for this part, each bin is represented with 12 bits to maintain the accuracy. As mentioned in [ 17 ], in the original HOG algorithm, the orientation and bin assignment module require W × H arctangent operations, W × H divisions, and 9 × W × H comparison operations. Although some previous works [ 10 ] use 18 × W × H multiplication operations instead of arctangent, by using our method, the bin assignment module does not use any multipliers.…”

“…We also perform division by shifting the histogram values, and therefore avoid a complex divider circuit. As mentioned in [ 17 ] in the original HOG algorithm, the block normalization step requires 9 × C multiplication (for the square of each histogram bin), addition and division operations, and B square root operations, where C is the total number of cells and B is the total number of blocks in an image window. Our simplification results in having 35 × B addition operations (for the adder tree) and 36 × B shifting operations (for four cells in each block).…”