Road sign recognition

Effect Introduction

In this section, we will use the road sign recognition model to recognize road sign images.

Note: Since the road sign images given during model training are fixed and the amount of data is limited, only the road signs in the routine's own images can be recognized, and real road signs may not be recognized.

Use CanMV IDE K230 to run the example source code, and you can see the camera image. When the camera is facing the road sign image, you can see the road sign framed and the road sign name displayed above the frame.

At the same time, the canmv IDE terminal will also print the corresponding name.

Routine source code

Copy the following source code to CanMV IDE K230 and run it.

xxxxxxxxxx
import os
import ujson
import aicube
from media.sensor import *
from media.display import *
from media.media import *
from time import *
import nncase_runtime as nn
import ulab.numpy as np
import time
import image
import gc
​
​
​
display_mode="lcd"
DISPLAY_WIDTH = ALIGN_UP(640, 16)
DISPLAY_HEIGHT = 480
​
OUT_RGB888P_WIDTH = ALIGN_UP(1080, 16)
OUT_RGB888P_HEIGH = 720
​
# 颜色盘 Color wheel
color_four = [(255, 220, 20, 60), (255, 119, 11, 32), (255, 0, 0, 142), (255, 0, 0, 230),
        (255, 106, 0, 228), (255, 0, 60, 100), (255, 0, 80, 100), (255, 0, 0, 70),
        (255, 0, 0, 192), (255, 250, 170, 30), (255, 100, 170, 30), (255, 220, 220, 0),
        (255, 175, 116, 175), (255, 250, 0, 30), (255, 165, 42, 42), (255, 255, 77, 255),
        (255, 0, 226, 252), (255, 182, 182, 255), (255, 0, 82, 0), (255, 120, 166, 157),
        (255, 110, 76, 0), (255, 174, 57, 255), (255, 199, 100, 0), (255, 72, 0, 118),
        (255, 255, 179, 240), (255, 0, 125, 92), (255, 209, 0, 151), (255, 188, 208, 182),
        (255, 0, 220, 176), (255, 255, 99, 164), (255, 92, 0, 73), (255, 133, 129, 255),
        (255, 78, 180, 255), (255, 0, 228, 0), (255, 174, 255, 243), (255, 45, 89, 255),
        (255, 134, 134, 103), (255, 145, 148, 174), (255, 255, 208, 186),
        (255, 197, 226, 255), (255, 171, 134, 1), (255, 109, 63, 54), (255, 207, 138, 255),
        (255, 151, 0, 95), (255, 9, 80, 61), (255, 84, 105, 51), (255, 74, 65, 105),
        (255, 166, 196, 102), (255, 208, 195, 210), (255, 255, 109, 65), (255, 0, 143, 149),
        (255, 179, 0, 194), (255, 209, 99, 106), (255, 5, 121, 0), (255, 227, 255, 205),
        (255, 147, 186, 208), (255, 153, 69, 1), (255, 3, 95, 161), (255, 163, 255, 0),
        (255, 119, 0, 170), (255, 0, 182, 199), (255, 0, 165, 120), (255, 183, 130, 88),
        (255, 95, 32, 0), (255, 130, 114, 135), (255, 110, 129, 133), (255, 166, 74, 118),
        (255, 219, 142, 185), (255, 79, 210, 114), (255, 178, 90, 62), (255, 65, 70, 15),
        (255, 127, 167, 115), (255, 59, 105, 106), (255, 142, 108, 45), (255, 196, 172, 0),
        (255, 95, 54, 80), (255, 128, 76, 255), (255, 201, 57, 1), (255, 246, 0, 122),
        (255, 191, 162, 208)]
​
root_path="/sdcard/mp_detect_guide/"
config_path=root_path+"deploy_config.json"
deploy_conf={}
debug_mode=0
​
class ScopedTiming:
    def __init__(self, info="", enable_profile=True):
        self.info = info
        self.enable_profile = enable_profile
​
    def __enter__(self):
        if self.enable_profile:
            self.start_time = time.time_ns()
        return self
​
    def __exit__(self, exc_type, exc_value, traceback):
        if self.enable_profile:
            elapsed_time = time.time_ns() - self.start_time
            print(f"{self.info} took {elapsed_time / 1000000:.2f} ms")
​
def read_deploy_config(config_path):
    # Open the JSON file to read deploy_config
    # 打开JSON文件以进行读取deploy_config
    with open(config_path, 'r') as json_file:
        try:
            # Load JSON data from a file
            # 从文件中加载JSON数据
            config = ujson.load(json_file)
            
            # Print data (other operations can be performed as needed)
            # 打印数据（可根据需要执行其他操作）
            #print(config)
        except ValueError as e:
            print("JSON 解析错误:", e)
    return config
​
def detection():
    print("det_infer start")
    # Use json to read content and initialize deployment variables
    # 使用json读取内容初始化部署变量
    deploy_conf=read_deploy_config(config_path)
    kmodel_name=deploy_conf["kmodel_path"]
    labels=deploy_conf["categories"]
    confidence_threshold= deploy_conf["confidence_threshold"]
    nms_threshold = deploy_conf["nms_threshold"]
    img_size=deploy_conf["img_size"]
    num_classes=deploy_conf["num_classes"]
    nms_option = deploy_conf["nms_option"]
    model_type = deploy_conf["model_type"]
    if model_type == "AnchorBaseDet":
        anchors = deploy_conf["anchors"][0] + deploy_conf["anchors"][1] + deploy_conf["anchors"][2]
    kmodel_frame_size = img_size
    frame_size = [OUT_RGB888P_WIDTH,OUT_RGB888P_HEIGH]
    strides = [8,16,32]
​
    # Calculate padding value
    # 计算padding值
    ori_w = OUT_RGB888P_WIDTH
    ori_h = OUT_RGB888P_HEIGH
    width = kmodel_frame_size[0]
    height = kmodel_frame_size[1]
    ratiow = float(width) / ori_w
    ratioh = float(height) / ori_h
    if ratiow < ratioh:
        ratio = ratiow
    else:
        ratio = ratioh
    new_w = int(ratio * ori_w)
    new_h = int(ratio * ori_h)
    dw = float(width - new_w) / 2
    dh = float(height - new_h) / 2
    top = int(round(dh - 0.1))
    bottom = int(round(dh + 0.1))
    left = int(round(dw - 0.1))
    right = int(round(dw - 0.1))
​
    # init kpu and load kmodel
    kpu = nn.kpu()
    ai2d = nn.ai2d()
    kpu.load_kmodel(root_path+kmodel_name)
    ai2d.set_dtype(nn.ai2d_format.NCHW_FMT,
                                   nn.ai2d_format.NCHW_FMT,
                                   np.uint8, np.uint8)
    ai2d.set_pad_param(True, [0,0,0,0,top,bottom,left,right], 0, [114,114,114])
    ai2d.set_resize_param(True, nn.interp_method.tf_bilinear, nn.interp_mode.half_pixel )
    ai2d_builder = ai2d.build([1,3,OUT_RGB888P_HEIGH,OUT_RGB888P_WIDTH], [1,3,height,width])
​
    # Initialize and configure the sensor
    # 初始化并配置sensor
    # sensor = Sensor(width=1280, height=960)
    sensor = Sensor()
    sensor.reset()
    # Set up mirroring
    # 设置镜像
    sensor.set_hmirror(False)
    # Set flip
    # 设置翻转
    sensor.set_vflip(False)
    # Channel 0 is directly given to the display VO, the format is YUV420
    # 通道0直接给到显示VO，格式为YUV420
    sensor.set_framesize(width = DISPLAY_WIDTH, height = DISPLAY_HEIGHT)
    sensor.set_pixformat(PIXEL_FORMAT_YUV_SEMIPLANAR_420)
    # Channel 2 is given to AI for algorithm processing, the format is RGB888
    # 通道2给到AI做算法处理，格式为RGB888
    sensor.set_framesize(width = OUT_RGB888P_WIDTH , height = OUT_RGB888P_HEIGH, chn=CAM_CHN_ID_2)
    sensor.set_pixformat(PIXEL_FORMAT_RGB_888_PLANAR, chn=CAM_CHN_ID_2)
    # Bind the output of channel 0 to vo
    # 绑定通道0的输出到vo
    sensor_bind_info = sensor.bind_info(x = 0, y = 0, chn = CAM_CHN_ID_0)
    Display.bind_layer(**sensor_bind_info, layer = Display.LAYER_VIDEO1)
    # Set to ST7701 display, default 640x480
    # 设置为ST7701显示，默认640x480
    Display.init(Display.ST7701, to_ide = True)
​
​
    #Create OSD image
    #创建OSD图像
    osd_img = image.Image(DISPLAY_WIDTH, DISPLAY_HEIGHT, image.ARGB8888)
    try:
        # media initialization
        # media初始化
        MediaManager.init()
        # Start the sensor
        # 启动sensor
        sensor.run()
        rgb888p_img = None
        ai2d_input_tensor = None
        data = np.ones((1,3,width,height),dtype=np.uint8)
        ai2d_output_tensor = nn.from_numpy(data)
        while  True:
            with ScopedTiming("total",debug_mode > 0):
                rgb888p_img = sensor.snapshot(chn=CAM_CHN_ID_2)
                # for rgb888planar
                if rgb888p_img.format() == image.RGBP888:
                    ai2d_input = rgb888p_img.to_numpy_ref()
                    ai2d_input_tensor = nn.from_numpy(ai2d_input)
                    ai2d_builder.run(ai2d_input_tensor, ai2d_output_tensor)
​
                    # set input
                    kpu.set_input_tensor(0, ai2d_output_tensor)
                    # run kmodel
                    kpu.run()
                    # get output
                    results = []
                    for i in range(kpu.outputs_size()):
                        out_data = kpu.get_output_tensor(i)
                        result = out_data.to_numpy()
                        result = result.reshape((result.shape[0]*result.shape[1]*result.shape[2]*result.shape[3]))
                        del out_data
                        results.append(result)
                    gc.collect()
​
                    # postprocess
                    if model_type == "AnchorBaseDet":
                        det_boxes = aicube.anchorbasedet_post_process( results[0], results[1], results[2], kmodel_frame_size, frame_size, strides, num_classes, confidence_threshold, nms_threshold, anchors, nms_option)
                    elif model_type == "GFLDet":
                        det_boxes = aicube.gfldet_post_process( results[0], results[1], results[2], kmodel_frame_size, frame_size, strides, num_classes, confidence_threshold, nms_threshold, nms_option)
                    else:
                        det_boxes = aicube.anchorfreedet_post_process( results[0], results[1], results[2], kmodel_frame_size, frame_size, strides, num_classes, confidence_threshold, nms_threshold, nms_option)
                    osd_img.clear()
                    if det_boxes:
                        for det_boxe in det_boxes:
                            x1, y1, x2, y2 = det_boxe[2],det_boxe[3],det_boxe[4],det_boxe[5]
                            w = float(x2 - x1) * DISPLAY_WIDTH // OUT_RGB888P_WIDTH
                            h = float(y2 - y1) * DISPLAY_HEIGHT // OUT_RGB888P_HEIGH
                            osd_img.draw_rectangle(int(x1 * DISPLAY_WIDTH // OUT_RGB888P_WIDTH) , int(y1 * DISPLAY_HEIGHT // OUT_RGB888P_HEIGH) , int(w) , int(h) , color=color_four[det_boxe[0]][1:])
                            label = labels[det_boxe[0]]
                            print("label:", label, det_boxe[0])
​
                            score = str(round(det_boxe[1],2))
                            osd_img.draw_string_advanced( int(x1 * DISPLAY_WIDTH // OUT_RGB888P_WIDTH) , int(y1 * DISPLAY_HEIGHT // OUT_RGB888P_HEIGH)-50,32, label + " " + score , color=color_four[det_boxe[0]][1:])
                    Display.show_image(osd_img, 0, 0, Display.LAYER_OSD3)
                    gc.collect()
                rgb888p_img = None
    except Exception as e:
        print(f"An error occurred during buffer used: {e}")
    finally:
        os.exitpoint(os.EXITPOINT_ENABLE_SLEEP)
        del ai2d_input_tensor
        del ai2d_output_tensor
        #Stop camera output
        #停止摄像头输出
        sensor.stop()
        # Initialize the display device
        #去初始化显示设备
        Display.deinit()
        #Release the media buffer
        #释放媒体缓冲区
        MediaManager.deinit()
        gc.collect()
        time.sleep(1)
        nn.shrink_memory_pool()
    print("det_infer end")
    return 0
​
​
if __name__=="__main__":
    detection()

Code flow

Main working principle:

1. Get live video stream from camera
2. Preprocess each frame and resize it to fit the model input requirements
3. Feed the processed image into the neural network for inference
4. Filter classification results based on confidence threshold
5. Draw the recognition results on the display interface
6. Loop to process the next frame

Code structure

1. Import dependencies

xxxxxxxxxx
import os
import ujson
import aicube
from media.sensor import *
from media.display import *
from media.media import *
from time import *
import nncase_runtime as nn
import ulab.numpy as np
import time
import image
import gc

2. Configuration initialization

xxxxxxxxxx
root_path="/sdcard/mp_detect_guide/"
config_path=root_path+"deploy_config.json"
deploy_conf={}
debug_mode=0

Among them, root_path specifies the path of the json parameter and model file.

3. Read json parameters

xxxxxxxxxx
def read_deploy_config(config_path):
    # Open the JSON file to read deploy_config
    # 打开JSON文件以进行读取deploy_config
    with open(config_path, 'r') as json_file:
        try:
            # Load JSON data from a file
            # 从文件中加载JSON数据
            config = ujson.load(json_file)
            
            # Print data (other operations can be performed as needed)
            # 打印数据（可根据需要执行其他操作）
            #print(config)
        except ValueError as e:
            print("JSON 解析错误:", e)
    return config

4. Main function

Initialize peripherals such as the camera and display, then read the camera image, pass it to the KPU for calculation, match it with the model file, output the recognition result, and then draw the recognition result on the display.

xxxxxxxxxx
def detection():
    print("det_infer start")
    # Use json to read content and initialize deployment variables
    # 使用json读取内容初始化部署变量
    deploy_conf=read_deploy_config(config_path)
    kmodel_name=deploy_conf["kmodel_path"]
    labels=deploy_conf["categories"]
    confidence_threshold= deploy_conf["confidence_threshold"]
    nms_threshold = deploy_conf["nms_threshold"]
    img_size=deploy_conf["img_size"]
    num_classes=deploy_conf["num_classes"]
    nms_option = deploy_conf["nms_option"]
    model_type = deploy_conf["model_type"]
    if model_type == "AnchorBaseDet":
        anchors = deploy_conf["anchors"][0] + deploy_conf["anchors"][1] + deploy_conf["anchors"][2]
    kmodel_frame_size = img_size
    frame_size = [OUT_RGB888P_WIDTH,OUT_RGB888P_HEIGH]
    strides = [8,16,32]
​
    # Calculate padding value
    # 计算padding值
    ori_w = OUT_RGB888P_WIDTH
    ori_h = OUT_RGB888P_HEIGH
    width = kmodel_frame_size[0]
    height = kmodel_frame_size[1]
    ratiow = float(width) / ori_w
    ratioh = float(height) / ori_h
    if ratiow < ratioh:
        ratio = ratiow
    else:
        ratio = ratioh
    new_w = int(ratio * ori_w)
    new_h = int(ratio * ori_h)
    dw = float(width - new_w) / 2
    dh = float(height - new_h) / 2
    top = int(round(dh - 0.1))
    bottom = int(round(dh + 0.1))
    left = int(round(dw - 0.1))
    right = int(round(dw - 0.1))
​
    # init kpu and load kmodel
    kpu = nn.kpu()
    ai2d = nn.ai2d()
    kpu.load_kmodel(root_path+kmodel_name)
    ai2d.set_dtype(nn.ai2d_format.NCHW_FMT,
                                   nn.ai2d_format.NCHW_FMT,
                                   np.uint8, np.uint8)
    ai2d.set_pad_param(True, [0,0,0,0,top,bottom,left,right], 0, [114,114,114])
    ai2d.set_resize_param(True, nn.interp_method.tf_bilinear, nn.interp_mode.half_pixel )
    ai2d_builder = ai2d.build([1,3,OUT_RGB888P_HEIGH,OUT_RGB888P_WIDTH], [1,3,height,width])
​
    # Initialize and configure the sensor
    # 初始化并配置sensor
    # sensor = Sensor(width=1280, height=960)
    sensor = Sensor()
    sensor.reset()
    # Set up mirroring
    # 设置镜像
    sensor.set_hmirror(False)
    # Set flip
    # 设置翻转
    sensor.set_vflip(False)
    # Channel 0 is directly given to the display VO, the format is YUV420
    # 通道0直接给到显示VO，格式为YUV420
    sensor.set_framesize(width = DISPLAY_WIDTH, height = DISPLAY_HEIGHT)
    sensor.set_pixformat(PIXEL_FORMAT_YUV_SEMIPLANAR_420)
    # Channel 2 is given to AI for algorithm processing, the format is RGB888
    # 通道2给到AI做算法处理，格式为RGB888
    sensor.set_framesize(width = OUT_RGB888P_WIDTH , height = OUT_RGB888P_HEIGH, chn=CAM_CHN_ID_2)
    sensor.set_pixformat(PIXEL_FORMAT_RGB_888_PLANAR, chn=CAM_CHN_ID_2)
    # Bind the output of channel 0 to vo
    # 绑定通道0的输出到vo
    sensor_bind_info = sensor.bind_info(x = 0, y = 0, chn = CAM_CHN_ID_0)
    Display.bind_layer(**sensor_bind_info, layer = Display.LAYER_VIDEO1)
    # Set to ST7701 display, default 640x480
    # 设置为ST7701显示，默认640x480
    Display.init(Display.ST7701, to_ide = True)
​
​
    #Create OSD image
    #创建OSD图像
    osd_img = image.Image(DISPLAY_WIDTH, DISPLAY_HEIGHT, image.ARGB8888)
    try:
        # media initialization
        # media初始化
        MediaManager.init()
        # Start the sensor
        # 启动sensor
        sensor.run()
        rgb888p_img = None
        ai2d_input_tensor = None
        data = np.ones((1,3,width,height),dtype=np.uint8)
        ai2d_output_tensor = nn.from_numpy(data)
        while  True:
            with ScopedTiming("total",debug_mode > 0):
                rgb888p_img = sensor.snapshot(chn=CAM_CHN_ID_2)
                # for rgb888planar
                if rgb888p_img.format() == image.RGBP888:
                    ai2d_input = rgb888p_img.to_numpy_ref()
                    ai2d_input_tensor = nn.from_numpy(ai2d_input)
                    ai2d_builder.run(ai2d_input_tensor, ai2d_output_tensor)
​
                    # set input
                    kpu.set_input_tensor(0, ai2d_output_tensor)
                    # run kmodel
                    kpu.run()
                    # get output
                    results = []
                    for i in range(kpu.outputs_size()):
                        out_data = kpu.get_output_tensor(i)
                        result = out_data.to_numpy()
                        result = result.reshape((result.shape[0]*result.shape[1]*result.shape[2]*result.shape[3]))
                        del out_data
                        results.append(result)
                    gc.collect()
​
                    # postprocess
                    if model_type == "AnchorBaseDet":
                        det_boxes = aicube.anchorbasedet_post_process( results[0], results[1], results[2], kmodel_frame_size, frame_size, strides, num_classes, confidence_threshold, nms_threshold, anchors, nms_option)
                    elif model_type == "GFLDet":
                        det_boxes = aicube.gfldet_post_process( results[0], results[1], results[2], kmodel_frame_size, frame_size, strides, num_classes, confidence_threshold, nms_threshold, nms_option)
                    else:
                        det_boxes = aicube.anchorfreedet_post_process( results[0], results[1], results[2], kmodel_frame_size, frame_size, strides, num_classes, confidence_threshold, nms_threshold, nms_option)
                    osd_img.clear()
                    if det_boxes:
                        for det_boxe in det_boxes:
                            x1, y1, x2, y2 = det_boxe[2],det_boxe[3],det_boxe[4],det_boxe[5]
                            w = float(x2 - x1) * DISPLAY_WIDTH // OUT_RGB888P_WIDTH
                            h = float(y2 - y1) * DISPLAY_HEIGHT // OUT_RGB888P_HEIGH
                            osd_img.draw_rectangle(int(x1 * DISPLAY_WIDTH // OUT_RGB888P_WIDTH) , int(y1 * DISPLAY_HEIGHT // OUT_RGB888P_HEIGH) , int(w) , int(h) , color=color_four[det_boxe[0]][1:])
                            label = labels[det_boxe[0]]
                            print("label:", label, det_boxe[0])
​
                            score = str(round(det_boxe[1],2))
                            osd_img.draw_string_advanced( int(x1 * DISPLAY_WIDTH // OUT_RGB888P_WIDTH) , int(y1 * DISPLAY_HEIGHT // OUT_RGB888P_HEIGH)-50,32, label + " " + score , color=color_four[det_boxe[0]][1:])
                    Display.show_image(osd_img, 0, 0, Display.LAYER_OSD3)
                    gc.collect()
                rgb888p_img = None
    except Exception as e:
        print(f"An error occurred during buffer used: {e}")
    finally:
        os.exitpoint(os.EXITPOINT_ENABLE_SLEEP)
        del ai2d_input_tensor
        del ai2d_output_tensor
        #Stop camera output
        #停止摄像头输出
        sensor.stop()
        # Initialize the display device
        #去初始化显示设备
        Display.deinit()
        #Release the media buffer
        #释放媒体缓冲区
        MediaManager.deinit()
        gc.collect()
        time.sleep(1)
        nn.shrink_memory_pool()
    print("det_infer end")
    return 0

Road sign pictures

go_straight:

turn_right:

Honking:

pedestrian_crossing:

speed_limit:

stop:

children_crossing:

Parking-A:

parking-B: