Initial commit: SmartCar Framework v0.1 — 龙芯2K0300智能车开发框架\n\n- HAL: GPIO/PWM/Encoder/Framebuffer 驱动\n- Control: PID/IMU/Motor/Servo 控制\n- Vision: HSV双Otsu→4点标定IPM→洪泛填充→逐行搜线\n- Strategy: 三区前瞻偏差+速度策略\n- Debug: 文件热调参+LCD预览+cv截帧\n- Scheduler: 5ms timerfd+epoll 中央调度器

model/config.json

@@ -0,0 +1,33 @@
+{
+  "seed": 42,
+  "data_path": "preprocessed/dataset_al8.npz",
+  "save_dir": "checkpoints_heat_v4",
+  "val_split": 0.15,
+  "batch_size": 64,
+  "epochs": 60,
+  "lr": 0.003,
+  "weight_decay": 0.0001,
+  "img_w": 160,
+  "img_h": 120,
+  "num_classes": 3,
+  "out_h": 15,
+  "out_w": 20,
+  "stride": 8,
+  "size_loss_weight": 2.0,
+  "bg_weight": 0.05,
+  "resume_from": "checkpoints_heat_v4/best.pth",
+  "ref_sizes": {
+    "0": [
+      75.35844421386719,
+      62.14570236206055
+    ],
+    "1": [
+      49.79265213012695,
+      38.633811950683594
+    ],
+    "2": [
+      67.58988189697266,
+      34.93575668334961
+    ]
+  }
+}

model/deploy_README.md

@@ -0,0 +1,39 @@
+# NanoDetHeat v4 部署说明
+
+## 模型文件
+
+| 文件 | 说明 |
+|------|------|
+| model_heat_v4.py | 模型架构定义 (PyTorch) |
+| best.pth | 训练权重 (141KB, FP32) |
+| model_heat_v4.onnx | ONNX 格式模型 |
+| config.json | 配置 + 类别参考尺寸 |
+
+## 输入
+
+- 格式: RGB, 160×120, 归一化到 [0,1]
+- 预处理: BGR→RGB 翻转, /255.0
+
+## 输出
+
+- 形状: [1, 6, 15, 20]
+- 通道 0-3: 类别 logits (红绿灯/锥桶/人行道/背景)
+- 通道 4-5: 尺寸预测 (宽/高, 需乘参考尺寸)
+
+## 后处理
+
+1. softmax 通道 0-3 → 概率
+2. 对每类找局部峰值 (3×3 邻域抑制)
+3. 筛选置信度 > 阈值 (推荐 0.8)
+4. 尺寸解码: 实际宽 = 预测宽 × 参考宽
+5. 中心坐标: cx=(gx+0.5)×8, cy=(gy+0.5)×8
+
+## 类别参考尺寸 (像素, 160×120)
+
+红绿灯: 73.7 × 60.9
+锥桶:   49.4 × 38.6
+人行道: 71.2 × 36.3
+
+## 推荐阈值
+
+th=0.8: 精确度 62.9%, 召回 88.7%

model/model.cpp

@@ -0,0 +1,13 @@
+#include "model.hpp"
+
+bool model_init()
+{
+    // TODO: 加载 TFLite 模型，初始化解释器
+    return false;
+}
+
+void model_infer(const uint8* image, int w, int h, float* output, int num_classes)
+{
+    // TODO: 运行推理，填充 output[0..num_classes-1]
+    for (int i = 0; i < num_classes; ++i) output[i] = 0.0f;
+}

model/model.hpp

@@ -0,0 +1,11 @@
+#pragma once
+#include "types.hpp"
+
+// TFLM 模型推理接口 — 占位实现，后续接入模型
+// 用法:
+//   1. 将 .tflite 模型转为 C header 数组放入 model/ 目录
+//   2. 实现 model_init() / model_infer()
+//   3. 在 element.cpp 中调用
+
+bool model_init();
+void model_infer(const uint8* image, int w, int h, float* output, int num_classes);

model/model_heat_v4.py

@@ -0,0 +1,100 @@
+"""
+NanoDetHeat v4: 残差 + SE注意力 + 扩通道
+MACs: ~8.7M, 参数: ~18K, 目标: 20 TPS @ 0.44 GOPS (79%)
+"""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+
+
+def conv_bn_relu(in_c, out_c, kernel, stride=1, groups=1):
+    pad = kernel // 2
+    return nn.Sequential(
+        nn.Conv2d(in_c, out_c, kernel, stride, pad, groups=groups, bias=False),
+        nn.BatchNorm2d(out_c),
+        nn.ReLU(inplace=True),
+    )
+
+
+class SEResDWBlock(nn.Module):
+    """残差DW + SE通道注意力"""
+    def __init__(self, in_c, out_c, stride=1, reduction=4):
+        super().__init__()
+        self.dw = conv_bn_relu(in_c, in_c, 3, stride=stride, groups=in_c)
+        self.pw = conv_bn_relu(in_c, out_c, 1)
+        
+        self.skip = nn.Identity() if (stride == 1 and in_c == out_c) else nn.Sequential(
+            nn.Conv2d(in_c, out_c, 1, stride, bias=False),
+            nn.BatchNorm2d(out_c),
+        )
+        
+        self.se = nn.Sequential(
+            nn.AdaptiveAvgPool2d(1),
+            nn.Conv2d(out_c, out_c // reduction, 1),
+            nn.ReLU(inplace=True),
+            nn.Conv2d(out_c // reduction, out_c, 1),
+            nn.Sigmoid(),
+        )
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        out = self.pw(self.dw(x))
+        out = out + self.skip(x)
+        return self.relu(out * self.se(out))
+
+
+class NanoDetHeatV4(nn.Module):
+    def __init__(self, num_classes=3):
+        super().__init__()
+        self.stem   = conv_bn_relu(3, 8, 3, stride=2)
+        self.block1 = SEResDWBlock(8, 16, stride=1)
+        self.block2 = SEResDWBlock(16, 28, stride=2)
+        self.block3 = SEResDWBlock(28, 40, stride=1)
+        self.block4 = SEResDWBlock(40, 56, stride=2)
+        self.block5 = SEResDWBlock(56, 64, stride=1)
+
+        self.shared = conv_bn_relu(64, 24, 1)
+        self.cls_head = nn.Conv2d(24, num_classes + 1, 1, bias=True)
+        self.size_head = nn.Conv2d(24, 2, 1, bias=True)
+        self._init_weights()
+
+    def _init_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+                if m.bias is not None: nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.constant_(m.weight, 1); nn.init.constant_(m.bias, 0)
+        self.cls_head.bias.data[-1] = 0.5  # 轻微背景偏置,焦点损失自动调整
+
+    def forward(self, x):
+        x = self.stem(x); x = self.block1(x); x = self.block2(x)
+        x = self.block3(x); x = self.block4(x); x = self.block5(x)
+        x = self.shared(x)
+        return torch.cat([self.cls_head(x), self.size_head(x)], dim=1)
+
+
+OUT_H, OUT_W = 15, 20
+STRIDE = 8
+
+
+if __name__ == "__main__":
+    m = NanoDetHeatV4()
+    p = sum(p.numel() for p in m.parameters())
+    
+    macs = (3*3*3*8*80*60 +                                        # stem
+            8*9*80*60 + 8*16*80*60 + 8*16*80*60 +                 # b1
+            16*9*40*30 + 16*28*40*30 + 16*28*40*30 +              # b2
+            28*9*40*30 + 28*40*40*30 +                             # b3
+            40*9*20*15 + 40*56*20*15 + 40*56*20*15 +              # b4
+            56*9*20*15 + 56*64*20*15 +                             # b5
+            64*24*20*15 + 24*4*20*15 + 24*2*20*15)                # heads
+    
+    print(f"Params: {p:,} = ~{p/1000:.1f}K")
+    print(f"MACs: {macs:,} = ~{macs/1e6:.2f}M")
+    print(f"20 TPS: {macs*20*2/1e9:.2f} GOPS ({(macs*20*2/0.44e9)*100:.0f}%)")
+    print(f"25 TPS: {macs*25*2/1e9:.2f} GOPS ({(macs*25*2/0.44e9)*100:.0f}%)")
+    x = torch.randn(1,3,120,160); y = m(x)
+    print(f"Output: {y.shape}")