HOG Circuit Design And Optimization For Pedestrian Detection

Pedestrian detection is a key technology in many application fields,such as surveillance,automobile system and robots.With the increasing complexity of the application environment,the response time of pedestrian detection is required to be higher and higher.Histograms of Oriented Gradients(HOG)is widely used to extract feature in pedestrian detection,because of its robustness and high detection accuracy for various texture objects.However,its high complexity and computation-intensive are the constraints of real-time detection of embedded systems.Therefore,this thesis mainly studies the optimization design of HOG algorithm and its hardware acceleration circuit,so as to to achieve high throughput HOG circuit and accelerate the speed of pedestrian detection.Firstly,the influence of calculation accuracy on the accuracy of HOG feature extraction is analyzed,and then18-bit fixed-point calculation is determined to realize HOG Circuit.According to the integration of bilinear interpolation weights and direction intervals in the iterative process of coordinate Rotation Digital Computer(CORDIC),the calculation process of CORDIC algorithm is improved to reduce the calculation time of histogram generation module.On this basis,the HOG hardware acceleration circuit is designed and optimized,which is included three parts: unit storage structure,histogram generation circuit and histogram normalization circuit.For cell storage structure,memory bandwidth is reduced by sharing and reusing cells,and efficient memory access mode is adopted to provide the memory area needed for synchronous access to different functional blocks in order to avoid repeated computation at different computing stages.For the histogram generation circuit,cascaded RAM is used to calculate the region histogram to optimize the circuit design of CORDIC algorithm.Finally,the cell histogram is synthesized,which improves the parallelism of the design.For the histogram normalization circuit,the polynomial approximation algorithm realized by fixed points can optimize the circuit structure of Newton iteration initial value calculation.Therefore,the complexity of the reciprocal square root calculation circuit is reduced and the running speed is improved at the same time.This thesis uses Zynq 7000 to verify the function and performance of HOG acceleration circuit.On the INRIA data set,90.65% of the detection accuracy is achieved.With the processing speed of 40 MHz clock frequency,the detection time of 12.1ms can be achieved with high accuracy.Based on the optimized hardware acceleration circuit of HOG and OV7670 camera,A pedestrian detection system is designed in this thesis,which achieves stable and accurate real-time pedestrian detection under 640×480@30fps.