Object Recognition on Raspberry Pi Zero2W

Project Introduction

Preparation: Includes the necessary Python environment and software package installation and deployment;

ONNX Model: Implements object recognition using lightweight ONNX models such as YOLOv5n;

Project Code: Includes flowcharts, key inference and execution code;

Demonstration: Showcases object recognition results in different scenarios.

ONNX

Open Neural Network Exchange (ONNX) is an open ecosystem that supports AI developers in choosing the right tools as their projects grow.

ONNX provides an open-source format for AI models, including deep learning and traditional ML.

It defines a scalable computational graph model, along with built-in operators and standard defined data types.

See: onnx/onnx: Open standard for machine learning interoperability.

YOLO

YOLO (You Only Look Once) is a popular object detection algorithm, renowned for its speed and accuracy. The core idea of ​​the YOLO algorithm is to transform the object detection problem into a regression problem, predicting bounding boxes and class probabilities directly from the input image using a single convolutional neural network (CNN).

YOLO's main advantages are its speed, enabling real-time processing of video streams. It also possesses strong versatility, allowing for transfer to new domains.

YOLOv5n

YOLOv5n represents an advancement in object detection methods. Based on the YOLOv5 model architecture developed by Ultralytics, it refines and optimizes the model architecture to achieve a higher accuracy-speed tradeoff in object detection tasks.

YOLOv5 provides an efficient solution for those seeking robust solutions in research and practical applications.

Preparation

For system installation and environment setup details, please refer to: [Raspberry Pi Zero2W] Introduction, System Installation, Face Detection.

Hardware Connection

If using SSH remote login, only a power supply connection is required;

If using local login, an HDMI video streaming cable, USB keyboard cable, etc., are required.

Library Installation

Execute the command `sudo apt install python3-opencv` to install OpenCV.

Install the `onnxruntime` library required for parsing ONNX models. Execute the command in the terminal.

sudo apt install python3-pip
sudo pip3 install onnxruntime --break-system-packages

If you encounter any problems, you can download and install the corresponding *.whl file according to your Linux system version and Python version.

wget https://pypi.tuna.tsinghua.edu.cn/packages/1c/a1/428ee29c6eaf09a6f6be56f836213f104618fb35ac6cc586ff0f477263eb/onnxruntime-1.23.2-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl
sudo pip install onnxruntime-1.23.2-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl --break-system-packages

Use `uname -a` to query Linux system information and `python --version` to check the Python version.

Model Download

Download the model file to your local machine (`./model`).

mkdir model
cd model
wget https://github.com/ultralytics/yolov5/releases/download/v7.0/yolov5n.onnx

Flowchart

Local board-side inference for object recognition is implemented using the lightweight onNX model of YOLOv5n, and the recognition results are displayed in a pop-up window.

Project Code

Create a new file using the command `touch or_onnx.py`, and add the following code using the nano text editor.

#!/usr/bin/env python3
import cv2, numpy as np, onnxruntime as ort
from pathlib import Path

# ---------- 命令行参数 ----------
parser = argparse.ArgumentParser(description='YOLOv5n-ONNX 物体识别')
parser.add_argument('image', nargs='?', default='./img/road.jpg',
help='待检测图片路径，缺省为 ./img/road.jpg')
args = parser.parse_args()
IMG_PATH = args.image
# ---------- 物体类别 ------------
COCO_NAMES = ["person","bicycle","car","motorcycle","airplane","bus","train","truck","boat","traffic light",
"fire hydrant","stop sign","parking meter","bench","bird","cat","dog","horse","sheep","cow",
"elephant","bear","zebra","giraffe","backpack","umbrella","handbag","tie","suitcase","frisbee",
"skis","snowboard","sports ball","kite","baseball bat","baseball glove","skateboard","surfboard",
"tennis racket","bottle","wine glass","cup","fork","knife","spoon","bowl","banana","apple",
"sandwich","orange","broccoli","carrot","hot dog","pizza","donut","cake","chair","couch",
"potted plant","bed","dining table","toilet","tv","laptop","mouse","remote","keyboard","cell phone",
"microwave","oven","toaster","sink","refrigerator","book","clock","vase","scissors","teddy bear",
"hair drier","toothbrush"]
# -------- 标签颜色 ------------
def generate_distinct_colors(num_colors):
colors = []

# 使用预定义的明显不同的颜色
base_colors = [
(255, 0, 0),    # 红色
(0, 255, 0),    # 绿色
(0, 0, 255),    # 蓝色
(255, 255, 0),  # 黄色
(255, 0, 255),  # 紫色
(0, 255, 255),  # 青色
(255, 128, 0),  # 橙色
(128, 0, 255),  # 深紫色
(0, 255, 128),  # 青绿色
(128, 255, 0),  # 黄绿色
(255, 0, 128),  # 粉红色
(0, 128, 255),  # 天蓝色
]

# 为常见类别设置固定颜色
fixed_colors = {
'person': (255, 0, 0),       # 红色
'car': (0, 255, 0),          # 绿色
'truck': (255, 255, 0),      # 黄色
'bus': (255, 128, 0),        # 橙色
'bicycle': (255, 0, 255),    # 紫色
'motorcycle': (128, 0, 255), # 深紫色
'traffic light': (0, 0, 255), # 蓝色
'stop sign': (0, 0, 255),    # 蓝色
'parking meter': (0, 255, 255), # 青色
}

# 生成所有类别的颜色
for i, name in enumerate(COCO_NAMES):
if name in fixed_colors:
colors.append(fixed_colors[name])
else:
color_idx = (i - len(fixed_colors)) % len(base_colors)
colors.append(base_colors[color_idx])

return colors
# --------- 文字颜色 ------------
def get_contrast_text_color(background_color):
"""
根据背景色返回对比度最高的文字颜色（黑色或白色）
使用相对亮度公式计算
"""
r, g, b = background_color

# 计算相对亮度（ITU-R BT.709标准）
luminance = 0.2126 * r + 0.7152 * g + 0.0722 * b

# 如果亮度大于128，使用黑色文字，否则使用白色文字
return (0, 0, 0) if luminance > 128 else (255, 255, 255)

CLASS_COLORS = generate_distinct_colors(len(COCO_NAMES))

MODEL_PATH = "model/yolov5n.onnx"
#IMG_PATH   = "img/road.jpg"

# ---------- 1. 加载模型 ----------
sess = ort.InferenceSession(MODEL_PATH, providers=['CPUExecutionProvider'])
in_name  = sess.get_inputs()[0].name
out_name = sess.get_outputs()[0].name

print(f"模型输入类型: {sess.get_inputs()[0].type}")
print(f"模型输出类型: {sess.get_outputs()[0].type}")

# ---------- 2. 读图 ----------
img0 = cv2.imread(IMG_PATH)
if img0 is None:
raise FileNotFoundError(f"无法找到图片: {IMG_PATH}")

# ---------- 3. 预处理 ----------
blob = cv2.resize(img0, (640, 640))
blob = blob.astype(np.float16) / 255.0
blob = blob.transpose(2, 0, 1)[None]

print(f"输入数据形状: {blob.shape}, 数据类型: {blob.dtype}")

# ---------- 4. 推理 ----------
outputs = sess.run([out_name], {in_name: blob})[0]
outputs = np.squeeze(outputs)

print(f"输出形状: {outputs.shape}")
print(f"输出范围: min={np.min(outputs):.6f}, max={np.max(outputs):.6f}")

if outputs.dtype == np.float16:
outputs = outputs.astype(np.float32)

# ---------- 5. 后处理 ----------
boxes, scores, class_ids = [], [], []
valid_detections = 0

for pred in outputs:
pred = pred.astype(np.float32)
x, y, w, h, conf = pred[:5]

if conf < 0.25:
continue

if not all(np.isfinite([x, y, w, h, conf])):
continue

if w <= 1 or h <= 1:
continue

class_probs = pred[5:]
cls_id = np.argmax(class_probs)
cls_conf = class_probs[cls_id]
final_conf = conf * cls_conf

# 坐标转换
if x > 1.0 or y > 1.0:
x_rel, y_rel, w_rel, h_rel = x/640.0, y/640.0, w/640.0, h/640.0
else:
x_rel, y_rel, w_rel, h_rel = x, y, w, h

valid_detections += 1
scores.append(float(final_conf))
class_ids.append(int(cls_id))

# 映射回原图坐标
x_scale = img0.shape[1] / 640
y_scale = img0.shape[0] / 640

x1 = int((x_rel - w_rel / 2) * img0.shape[1])
y1 = int((y_rel - h_rel / 2) * img0.shape[0])
x2 = int((x_rel + w_rel / 2) * img0.shape[1])
y2 = int((y_rel + h_rel / 2) * img0.shape[0])

x1 = max(0, min(x1, img0.shape[1]))
y1 = max(0, min(y1, img0.shape[0]))
x2 = max(0, min(x2, img0.shape[1]))
y2 = max(0, min(y2, img0.shape[0]))

width = x2 - x1
height = y2 - y1

if width > 0 and height > 0:
boxes.append([x1, y1, width, height])

print(f"总检测数: {len(outputs)}, 有效检测: {valid_detections}, 最终边界框: {len(boxes)}")

# ---------- 6. NMS和绘制结果 ----------
if boxes:
indices = cv2.dnn.NMSBoxes(boxes, scores, score_threshold=0.25, nms_threshold=0.45)

if indices is not None and len(indices) > 0:
print(f"NMS后保留 {len(indices)} 个检测结果")

# 存储检测结果用于排序输出
detection_results = []
for i in indices.flatten():
x, y, w, h = boxes[i]
cls_id = class_ids[i]
score = scores[i]
class_name = COCO_NAMES[cls_id]
detection_results.append((class_name, score, [x, y, w, h]))

# 获取类别颜色
bg_color = CLASS_COLORS[cls_id]

# 计算对比文字颜色
text_color = get_contrast_text_color(bg_color)

# 绘制边界框（细线）
cv2.rectangle(img0, (x, y), (x + w, y + h), bg_color, 1)

# 创建标签文本
label = f"{class_name} {score:.2f}"

# 计算文本尺寸
font = cv2.FONT_HERSHEY_SIMPLEX
font_scale = 0.5
thickness = 1
(text_width, text_height), baseline = cv2.getTextSize(label, font, font_scale, thickness)

# 标签背景位置
text_x = x
text_y = y - 5 if y - 5 > text_height else y + text_height + 5

# 确保标签不超出图像上边界
if text_y - text_height - baseline < 0:
text_y = y + text_height + 5

# 绘制标签背景（填充矩形）
cv2.rectangle(img0,
(text_x, text_y - text_height - baseline),
(text_x + text_width, text_y + baseline),
bg_color, -1)

# 绘制标签文本（自适应颜色）
cv2.putText(img0, label, (text_x, text_y), font, font_scale,
text_color, thickness, cv2.LINE_AA)

# 按置信度排序并输出
detection_results.sort(key=lambda x: x[1], reverse=True)
for class_name, score, bbox in detection_results:
x, y, w, h = bbox
print(f"检测到: {class_name} ({score:.3f}) at [{x}, {y}, {w}, {h}]")

else:
print("NMS后没有保留任何检测结果")
else:
print("没有有效的检测边界框")

# ---------- 7. 显示结果 ----------
cv2.imshow("YOLOv5n_ONNX", img0)
cv2.waitKey(0)
cv2.destroyAllWindows()

Save the code.

Demonstration:

Execute `python or_onnx.py./img/road.jpg` in the terminal. A pop-up window will display the object recognition results.

At the same time, the terminal prints information.

More scenarios

Thanks to GeKenflex for providing the funds to support.