基于华为atlas的重车（满载）空车（空载）识别

该教程主要是想摸索出华为atlas的基于ACL的推理模式。最终实现通过煤矿磅道上方的摄像头，识别出车辆的重车（满载）、空车（空载）情况。本质上是一个简单的检测问题。

但是整体探索过程比较坎坷，Tianxiaomo的代码可以基于原始yolov4模型进行推理，可以转化onnx，但是训练过程我感觉代码有问题，loss很大，也没检测框输出。同时输出的结果的维度和atlas教程的维度也不一样。对于这2个问题，第1个问题训练效果不对选择使用原始darknet网络解决，第2个问题输出维度和atlas不一样通过重新实现atlas后处理代码实现。

风雨兼程，终见彩虹；艰辛耕耘，方得硕果。

darknet数据集制作及配置文件修改：

（1）数据集采用labelimg工具标注为VOC格式，一共标注了1087张图片。

（2）数据集格式如下，

其中，VOC2025为我自己的数据集起的名字，你也可以起别的名字，Annotations存放XML文件，Main中存放，train.txt，val.txt，txt中只写图片的名字，一行一个。JPEGImages中存放图片。labels中存放由XML生成的txt文件。

（3）修改scripts下面的voc_label.py，将数据集的目录修改为自己的目录，

#开始几行
sets=[('2025', 'train'), ('2025', 'val')]
classes = ["full", "empty"]
#最后2行
os.system("cat 2025_train.txt 2025_val.txt > train.txt")
os.system("cat 2025_train.txt 2025_val.txt > train.all.txt")

然后执行

Python3 scripts/voc_label.py

就会生成labels文件夹，以及文件夹下面的txt标记，以及train.txt 和train.all.txt

其中,train.txt中存储路径+图片名，一行一个

/data/jxl/darknet/VOCdevkit/VOC2025/JPEGImages/3743_01467.jpg
/data/jxl/darknet/VOCdevkit/VOC2025/JPEGImages/3743_01468.jpg
/data/jxl/darknet/VOCdevkit/VOC2025/JPEGImages/3743_01469.jpg
/data/jxl/darknet/VOCdevkit/VOC2025/JPEGImages/3743_01559.jpg

Labels文件夹下每个图片对应一个txt文件，里面存储类别框坐标的归一化值

0 0.6794407894736842 0.5394736842105263 0.5516447368421052 0.9195906432748537

（4）修改，cfg/fullempty.data

classes= 2
train  = ./VOCdevkit/VOC2025/ImageSets/Main/train.txt
valid  = ./VOCdevkit/VOC2025/ImageSets/Main/val.txt
names = ./data/fullempty.names
backup = ./pjreddie/backup/

class为训练的类别数

train为训练集train.txt

valid为验证集val.txt

names为fullempty.names，里面为自己训练的目标名称

backup为weights的存储位置

（5）修改cfg/yolov4-fullempty.cfg

修改每个classes=2

修改最后一个卷基层，filters和最后一个region的classes，num参数是因为yolov4有3个分支，每个分支3个anchor。

其中，filters=num×（classes + coords + 1）=3*(2+4+1)=21，这里我有2个类别。

（6）修改data/fullempty.names

full
empty

darknet模型训练：

./darknet detector train ./cfg/fullempty.data ./cfg/yolov4-fullempty.cfg  ./yolov4.weights -clear

darknet的.weights模型测试：

./darknet detect  ./cfg/yolov4-fullempty.cfg  ./pjreddie/backup/yolov4-fullempty_last.weights  ./VOCdevkit/VOC2025/JPEGImages/2793_00847.jpg

Pytorch代码配置文件修改：

#cfg.py，

Cfg.use_darknet_cfg = True
Cfg.cfgfile = os.path.join(_BASE_DIR, 'cfg', 'yolov4-custom.cfg')

#cfg/yolov4-custom.cfg，

Pytorch代码bug修改:

#train.py211行，

pred_ious = bboxes_iou(pred[b].view(-1, 4), truth_box, xyxy=False)修改为，
pred_ious = bboxes_iou(pred[b].contiguous().view(-1, 4), truth_box, xyxy=False)

# dataset.py, get_image_id函数，因为我的图片命名规则是Id_id.jpg，所以将2个id拼接起来作为最终的id。

parts = filename.split('.')[0].split('_')
id = int(parts[0]+ parts[1])
return id

基于pytorch代码的.weights模型测试:

python3 demo.py -cfgfile ./cfg/yolov4-custom.cfg -weightfile ./yolov4-fullempty_last.weights -imgfile ./full_empty_dataset/images/2793_00847.jpg -torch False

.weights模型转onnx模型：

python3 demo_darknet2onnx.py ./cfg/yolov4-custom.cfg ./data/full_empty.names ./yolov4-fullempty_last.weights ./full_empty_dataset/images/2793_00847.jpg 1

onnx模型转om模型：

atc --model=./yolov4_1_3_608_608_static.onnx --framework=5 --output=yolov4_bs1 --input_shape="input:1,3,608, 608"  --soc_version=Ascend310P3 --input_format=NCHW

atlas推理代码编写：

#yolov4.py

import sys
sys.path.append("./common/acllite")
import os
import numpy as np
import acl
import cv2
import time
from acllite_model import AclLiteModel
from acllite_resource import AclLiteResourcefrom utils import post_processing, plot_boxes_cv2MODEL_PATH = "./model/yolov4_bs1.om"#ACL resource initialization
acl_resource = AclLiteResource()
acl_resource.init()
#load model
model = AclLiteModel(MODEL_PATH)class YOLOV4(object):def __init__(self):self.MODEL_PATH = MODEL_PATHself.MODEL_WIDTH = 608self.MODEL_HEIGHT = 608self.class_names= ['full', 'empty']self.model = modeldef preprocess(self, bgr_img):sized = cv2.resize(bgr_img.copy(), (self.MODEL_WIDTH, self.MODEL_HEIGHT))sized = cv2.cvtColor(sized, cv2.COLOR_BGR2RGB)new_image = sized.astype(np.float32)new_image = new_image / 255.0new_image = new_image.transpose(2, 0, 1).copy()return new_imagedef process(self, bgr_img):height, width = bgr_img.shape[:2]#preprocessdata = self.preprocess(bgr_img)#(3, 608, 608)#Send into model inferenceresult_list = self.model.execute([data,])    #Process inference resultsconf_thresh, nms_thresh = 0.4, 0.6boxes = post_processing(conf_thresh, nms_thresh, result_list, height, width)return boxesdef draw(self, bgr_img, boxes):drawed_img = plot_boxes_cv2(bgr_img, boxes[0], class_names=self.class_names)return drawed_imgdef test_image():yolov4 = YOLOV4()img_name = "./data/3553_00173.jpg"#read imagebgr_img = cv2.imread(img_name)t1 = time.time()boxes = yolov4.process(bgr_img)t2 = time.time()drawed_img = yolov4.draw(bgr_img, boxes)t3 = time.time()print("result = ", len(boxes[0]), boxes, t2-t1, t3-t2)cv2.imwrite("out.jpg", drawed_img)if __name__ == '__main__':test_image()

#utils.py

import sys
import os
import time
import math
import numpy as npimport itertools
import struct  # get_image_size
import imghdr  # get_image_sizedef sigmoid(x):return 1.0 / (np.exp(-x) + 1.)def softmax(x):x = np.exp(x - np.expand_dims(np.max(x, axis=1), axis=1))x = x / np.expand_dims(x.sum(axis=1), axis=1)return xdef bbox_iou(box1, box2, x1y1x2y2=True):if x1y1x2y2:mx = min(box1[0], box2[0])Mx = max(box1[2], box2[2])my = min(box1[1], box2[1])My = max(box1[3], box2[3])w1 = box1[2] - box1[0]h1 = box1[3] - box1[1]w2 = box2[2] - box2[0]h2 = box2[3] - box2[1]else:w1 = box1[2]h1 = box1[3]w2 = box2[2]h2 = box2[3]mx = min(box1[0], box2[0])Mx = max(box1[0] + w1, box2[0] + w2)my = min(box1[1], box2[1])My = max(box1[1] + h1, box2[1] + h2)uw = Mx - mxuh = My - mycw = w1 + w2 - uwch = h1 + h2 - uhcarea = 0if cw <= 0 or ch <= 0:return 0.0area1 = w1 * h1area2 = w2 * h2carea = cw * chuarea = area1 + area2 - careareturn carea / uareadef nms_cpu(boxes, confs, nms_thresh=0.5, min_mode=False):x1 = boxes[:, 0]y1 = boxes[:, 1]x2 = boxes[:, 2]y2 = boxes[:, 3]areas = (x2 - x1) * (y2 - y1)order = confs.argsort()[::-1]keep = []while order.size > 0:idx_self = order[0]idx_other = order[1:]keep.append(idx_self)xx1 = np.maximum(x1[idx_self], x1[idx_other])yy1 = np.maximum(y1[idx_self], y1[idx_other])xx2 = np.minimum(x2[idx_self], x2[idx_other])yy2 = np.minimum(y2[idx_self], y2[idx_other])w = np.maximum(0.0, xx2 - xx1)h = np.maximum(0.0, yy2 - yy1)inter = w * hif min_mode:over = inter / np.minimum(areas[order[0]], areas[order[1:]])else:over = inter / (areas[order[0]] + areas[order[1:]] - inter)inds = np.where(over <= nms_thresh)[0]order = order[inds + 1]return np.array(keep)def plot_boxes_cv2(img, boxes, class_names=None, color=None):import cv2img = np.copy(img)colors = np.array([[1, 0, 1], [0, 0, 1], [0, 1, 1], [0, 1, 0], [1, 1, 0], [1, 0, 0]], dtype=np.float32)def get_color(c, x, max_val):ratio = float(x) / max_val * 5i = int(math.floor(ratio))j = int(math.ceil(ratio))ratio = ratio - ir = (1 - ratio) * colors[i][c] + ratio * colors[j][c]return int(r * 255)width = img.shape[1]height = img.shape[0]for i in range(len(boxes)):box = boxes[i]x1 = int(box[0])y1 = int(box[1])x2 = int(box[2])y2 = int(box[3])bbox_thick = int(0.6 * (height + width) / 600)if color:rgb = colorelse:rgb = (255, 0, 0)if len(box) >= 7 and class_names:cls_conf = box[5]cls_id = box[6]print('%s: %f' % (class_names[cls_id], cls_conf))classes = len(class_names)offset = cls_id * 123457 % classesred = get_color(2, offset, classes)green = get_color(1, offset, classes)blue = get_color(0, offset, classes)if color is None:rgb = (red, green, blue)msg = str(class_names[cls_id])+" "+str(round(cls_conf,3))t_size = cv2.getTextSize(msg, 0, 0.7, thickness=bbox_thick // 2)[0]c1, c2 = (x1,y1), (x2, y2)c3 = (c1[0] + t_size[0], c1[1] - t_size[1] - 3)cv2.rectangle(img, (x1,y1), (np.int32(c3[0]), np.int32(c3[1])), rgb, -1)img = cv2.putText(img, msg, (c1[0], np.int32(c1[1] - 2)), cv2.FONT_HERSHEY_SIMPLEX,0.7, (0,0,0), bbox_thick//2,lineType=cv2.LINE_AA)#cv2.rectangle(img, (x1,y1), (np.float32(c3[0]), np.float32(c3[1])), rgb, -1)#img = cv2.putText(img, msg, (c1[0], np.float32(c1[1] - 2)), cv2.FONT_HERSHEY_SIMPLEX,0.7, (0,0,0), bbox_thick//2,lineType=cv2.LINE_AA)img = cv2.rectangle(img, (x1, y1), (x2, y2), rgb, bbox_thick)return imgdef read_truths(lab_path):if not os.path.exists(lab_path):return np.array([])if os.path.getsize(lab_path):truths = np.loadtxt(lab_path)truths = truths.reshape(truths.size / 5, 5)  # to avoid single truth problemreturn truthselse:return np.array([])def load_class_names(namesfile):class_names = []with open(namesfile, 'r') as fp:lines = fp.readlines()for line in lines:line = line.rstrip()class_names.append(line)return class_namesdef post_processing(conf_thresh, nms_thresh, output, height, width):# anchors = [12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401]# num_anchors = 9# anchor_masks = [[0, 1, 2], [3, 4, 5], [6, 7, 8]]# strides = [8, 16, 32]# anchor_step = len(anchors) // num_anchors# [batch, num, 1, 4]box_array = output[0]# [batch, num, num_classes]confs = output[1]t1 = time.time()if type(box_array).__name__ != 'ndarray':box_array = box_array.cpu().detach().numpy()confs = confs.cpu().detach().numpy()num_classes = confs.shape[2]# [batch, num, 4]box_array = box_array[:, :, 0]# [batch, num, num_classes] --> [batch, num]max_conf = np.max(confs, axis=2)max_id = np.argmax(confs, axis=2)t2 = time.time()bboxes_batch = []for i in range(box_array.shape[0]):argwhere = max_conf[i] > conf_threshl_box_array = box_array[i, argwhere, :]l_max_conf = max_conf[i, argwhere]l_max_id = max_id[i, argwhere]bboxes = []# nms for each classfor j in range(num_classes):cls_argwhere = l_max_id == jll_box_array = l_box_array[cls_argwhere, :]ll_max_conf = l_max_conf[cls_argwhere]ll_max_id = l_max_id[cls_argwhere]keep = nms_cpu(ll_box_array, ll_max_conf, nms_thresh)if (keep.size > 0):ll_box_array = ll_box_array[keep, :]ll_max_conf = ll_max_conf[keep]ll_max_id = ll_max_id[keep]for k in range(ll_box_array.shape[0]):bboxes.append([ll_box_array[k, 0]*width, ll_box_array[k, 1]*height, ll_box_array[k, 2]*width, ll_box_array[k, 3]*height, ll_max_conf[k], ll_max_conf[k], ll_max_id[k]])bboxes_batch.append(bboxes)t3 = time.time()return bboxes_batch

atlas推理代码测试：