CN114821660A - A pedestrian detection and reasoning method based on embedded devices - Google Patents

Abstract

The invention provides a pedestrian detection and reasoning method based on an embedded device, which is used for running a calculation-intensive pedestrian detection model on a low-power embedded device. The edge-end deep learning device based on RISC‑V architecture is adopted, the MCU development board is used as the hardware platform, the flat-head brother wujian100 open source IP is used as the MCU core, and the serial port, HDMI interface and OV5640 camera are onboard. Obtain the training data, and train the pedestrian detection model MobileNetV1‑SSD; calculate the quantization factor of the model weight; calculate the activation value quantization factor of each layer by minimizing the mean square error; quantize each operator in the model, convert the floating point Type model weights are quantized into int8 data types, and activation values are quantized into uint8 data types; model inference and inverse quantization, inverse quantization of inference results into int32 data types; compile on the MCU development board and run the model.

Description

A pedestrian detection and reasoning method based on embedded devices

technical field

The invention relates to a pedestrian detection and reasoning method based on an embedded device, and belongs to the technical field of pedestrian detection.

Background technique

In recent years, neural network models have been widely used in many fields and achieved very good results, especially in the field of pedestrian detection. However, due to the high model complexity and large model size of the pedestrian detection neural network model, the inference efficiency is low and the inference time is long, especially when running on low-performance mobile devices and low-power devices. Therefore, how to design a model with low resource consumption, which can be predicted in real time and at the same time guarantees the prediction accuracy has become a real problem. On low-power devices like MCUs, models with low resource consumption are required. In addition, many MCUs do not support floating-point arithmetic, which limits the application of models. Model quantization has achieved good results in dealing with these problems. Quantizing the model from floating-point type to fixed-point type can effectively reduce the model size, improve the model inference speed, and increase the supported embedded device types.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a pedestrian detection and reasoning method based on an embedded device, which ensures the accuracy of the model, and at the same time improves the reasoning speed of the model by calculating the quantization factor in advance.

The present invention is achieved by the following technical solutions in order to achieve the above object:

1. A pedestrian detection and reasoning method based on an embedded device is characterized in that, comprising the following steps:

1) Obtain training data and train a pedestrian detection model;

2) Calculate the quantization factor of the model weight, and calculate the model weight quantization factor based on the quantization range by calculating the maximum value of the absolute value of the model weight;

3) Calculate the activation value quantization factor of each layer by minimizing the mean square error. Based on some test data sets, calculate the mean square error of the quantized output and the unquantized output of each layer, and obtain the activation value quantization by minimizing the mean square error. factor;

4) Quantize each operator in the model, adopt asymmetric quantization, quantize the floating-point type model weight into int8 data type, and quantize the activation value into uint8 data type;

5) Model inference and inverse quantization, using the weights and activation values of the quantized fixed-point type for model inference, and inversely quantizing the inference results into int32 data types;

6) Compile and run the model on the MCU development board.

Preferably, the mean square error formula is as follows:

α=r/255

分别表示不量化输出以及量化后输出，量化范围r>0,α表示量化因子,clip指将激活值裁剪到[-r,r]范围，rounding是指将浮点数近似到最近的整数。where: y _i and

Represents the unquantized output and the quantized output, respectively, the quantization range r>0, α represents the quantization factor, clip refers to clipping the activation value to the [-r, r] range, and rounding refers to approximating the floating point number to the nearest integer.

Preferably, the pedestrian detection model adopts a lightweight network MobileNetV1-SSD.

Preferably, the quantization range of the quantization factor is [-128, 127].

The advantage of the present invention is that the present invention calculates the activation value quantization factor of each layer by minimizing the mean square error. The method ensures the accuracy of the model, and at the same time, by calculating the quantization factor in advance, the inference speed of the model is improved, and it can be applied to pedestrian detection and inference. In addition, running the pedestrian detection inference model on the MCU development board reduces the power consumption of the model.

Description of drawings

The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification, and are used to explain the present invention together with the embodiments of the present invention, and do not constitute a limitation to the present invention.

FIG. 1 is a schematic diagram of the flow structure of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

An embedded device-based pedestrian detection inference method for running computationally intensive pedestrian detection models onto low-power embedded devices. Using edge-end deep learning equipment based on RISC-V architecture, Xilinx MCU development board as hardware platform, Pingtou brother wujian100 open source IP as MCU core, onboard serial port, HDMI interface and OV5640 camera, real-time image data can be realized through the camera By running the pedestrian detection inference model on the MCU, the detection results are output to the peripheral devices through the rear serial port and HDMI.

1) Obtain the training data and train the pedestrian detection model. The model adopts the lightweight network MobileNetV1-SSD.

2) Calculate the quantization factor of the model weight. By calculating the maximum value of the absolute value of the model weight, the model weight quantization factor is calculated based on the quantization range. The model weight is quantized into int8 type, so the quantization range is [-128,127];

分别表示不量化输出以及量化后输出。量化范围r(r>0),量化因子α,clip指将激活值裁剪到[-r,r]范围，rounding是指将浮点数近似到最近的整数。3) Calculate the activation value quantization factor of each layer by minimizing the mean square error. Based on some test data sets, calculate the mean square error of the quantized output and the unquantized output of each layer, and obtain the activation value quantization by minimizing the mean square error. factor. The following is the mean square error formula, y _i and

Represent unquantized output and quantized output, respectively. Quantization range r (r>0), quantization factor α, clip refers to clipping the activation value to the [-r,r] range, and rounding refers to approximating floating point numbers to the nearest integer.

α=r/255

4) Perform quantization for each operator in the model, using asymmetric quantization to quantize the floating-point model weights into int8 data types, and quantize the activation values into uint8 data types.

5) Model inference and inverse quantization, use the quantized fixed-point type weight and activation value for model inference, and inversely quantize the inference result into an int32 data type, and the activation value quantization factor and inverse quantization factor participate in the calculation by shifting to avoid Use floating point calculations.

6) Compile on the MCU development board, run the model, preprocess the image data captured by the camera and transmit it to the model, and output the pedestrian detection result to the peripheral device through the rear serial port and HDMI.

Claims (4)

1. a pedestrian detection and reasoning method based on embedded equipment, is characterized in that, comprises the following steps: 1) Obtain training data and train a pedestrian detection model; 2) Calculate the quantization factor of the model weight, and calculate the model weight quantization factor based on the quantization range by calculating the maximum value of the absolute value of the model weight; 3) Calculate the activation value quantization factor of each layer by minimizing the mean square error. Based on some test data sets, calculate the mean square error of the quantized output and the unquantized output of each layer, and obtain the activation value quantization by minimizing the mean square error. factor; 4) Quantize each operator in the model, adopt asymmetric quantization, quantize the floating-point type model weight into int8 data type, and quantize the activation value into uint8 data type; 5) Model inference and inverse quantization, using the weights and activation values of the quantized fixed-point type for model inference, and inversely quantizing the inference results into int32 data types; 6) Compile and run the model on the MCU development board. 2. the pedestrian detection and reasoning method based on embedded device according to claim 1, is characterized in that, described mean square error formula is as follows:

α=r/255

where: y _i and

Represents the unquantized output and the quantized output, respectively, the quantization range r>0, α represents the quantization factor, clip refers to clipping the activation value to the [-r, r] range, and rounding refers to approximating the floating point number to the nearest integer. 3 . The pedestrian detection and reasoning method based on an embedded device according to claim 1 , wherein the pedestrian detection model adopts a lightweight network MobileNetV1-SSD. 4 . 4 . The pedestrian detection and reasoning method based on an embedded device according to claim 1 , wherein the quantization factor has a quantization range of [-128, 127]. 5 .

Images