ICCV 2025：从视线到运动预测自动驾驶视觉注意力

论文来源： ICCV 2025
会议： ICCV 2025
核心创新： 视线 → 车辆运动 → 预测未来视觉注意力

---

论文信息

项目	内容
标题	From Gaze to Movement: Predicting Visual Attention for Autonomous Driving
会议	ICCV 2025
链接	https://openaccess.thecvf.com/content/ICCV2025/papers/Huang_From_Gaze_to_Movement_Predicting_Visual_Attention_for_Autonomous_Driving_ICCV_2025_paper.pdf

---

核心问题：驾驶员视线预测

传统视线估计局限：

静态视线估计：
- 只估计当前视线方向
- 无法预测未来注意力焦点
- 未考虑车辆运动影响

Euro NCAP 2026 要求：
- 预测驾驶员即将看向的位置
- 提前检测分心趋势
- 与 ADAS 场景协同

---

方法创新：视线-运动联合建模

import torch
import torch.nn as nn
import torch.nn.functional as F

class GazeMovementPredictor(nn.Module):
    """
    视线-运动预测模型
    
    ICCV 2025 论文核心：
    基于当前视线 + 车辆运动 → 预测未来视觉注意力
    
    应用场景：
    1. 驾驶员注意力预测
    2. ADAS 警告时机优化
    3. 分心趋势检测
    """
    
    def __init__(self,
                 gaze_dim: int = 2,  # (pitch, yaw)
                 vehicle_dim: int = 6,  # (速度, 方向盘, 车道位置等)
                 hidden_dim: int = 128,
                 future_steps: int = 5):  # 预测未来 5 帧
        super().__init__()
        
        # 视线编码器
        self.gaze_encoder = nn.Sequential(
            nn.Linear(gaze_dim, hidden_dim),
            nn.ReLU(inplace=True),
            nn.Linear(hidden_dim, hidden_dim)
        )
        
        # 车辆运动编码器
        self.vehicle_encoder = nn.Sequential(
            nn.Linear(vehicle_dim, hidden_dim),
            nn.ReLU(inplace=True),
            nn.Linear(hidden_dim, hidden_dim)
        )
        
        # 融合层
        self.fusion = nn.Sequential(
            nn.Linear(hidden_dim * 2, hidden_dim),
            nn.ReLU(inplace=True),
            nn.Linear(hidden_dim, hidden_dim)
        )
        
        # 未来视线预测头
        self.future_gaze_head = nn.Linear(hidden_dim, gaze_dim * future_steps)
        
        # 注意力区域预测头
        self.attention_head = nn.Linear(hidden_dim, 3)  # (前方, 左侧, 右侧) 概率
        
        self.future_steps = future_steps
        self.gaze_dim = gaze_dim
    
    def forward(self, 
                current_gaze: torch.Tensor,
                vehicle_state: torch.Tensor) -> dict:
        """
        前向传播
        
        Args:
            current_gaze: 当前视线 (pitch, yaw), shape=(B, 2)
            vehicle_state: 车辆状态 [速度, 方向盘, 车道偏移等], shape=(B, 6)
        
        Returns:
            dict: {
                "future_gaze": 未来视线轨迹, shape=(B, future_steps, 2)
                "attention_probs": 注意力区域概率, shape=(B, 3)
                "predicted_deviation": 预测视线偏离, shape=(B, 1)
            }
        """
        # 编码
        gaze_features = self.gaze_encoder(current_gaze)
        vehicle_features = self.vehicle_encoder(vehicle_state)
        
        # 融合
        combined = torch.cat([gaze_features, vehicle_features], dim=1)
        fused_features = self.fusion(combined)
        
        # 未来视线预测
        future_gaze_flat = self.future_gaze_head(fused_features)
        future_gaze = future_gaze_flat.view(-1, self.future_steps, self.gaze_dim)
        
        # 注意力区域预测
        attention_probs = F.softmax(self.attention_head(fused_features), dim=1)
        
        # 预测偏离程度（基于未来轨迹）
        predicted_deviation = self._compute_deviation(future_gaze)
        
        return {
            "future_gaze": future_gaze,
            "attention_probs": attention_probs,
            "predicted_deviation": predicted_deviation
        }
    
    def _compute_deviation(self, future_gaze: torch.Tensor) -> torch.Tensor:
        """
        计算预测偏离程度
        
        如果未来视线持续偏离道路中心，预测偏离高
        """
        # 假设道路中心视线为 (0, 0)
        road_center = torch.zeros(self.gaze_dim)
        
        # 计算每帧偏离
        deviations = torch.norm(future_gaze - road_center, dim=2)
        
        # 平均偏离
        mean_deviation = deviations.mean(dim=1)
        
        return mean_deviation


# 测试示例
if __name__ == "__main__":
    model = GazeMovementPredictor()
    
    # 模拟输入
    current_gaze = torch.tensor([
        [0.1, 0.05],  # 正常：略微向右
        [0.5, -0.3],  # 分心：大角度偏离
    ])
    
    vehicle_state = torch.tensor([
        [60, 0.1, 0.02, 0, 0, 0],  # 正常驾驶
        [60, 0.5, 0.2, 0, 0, 0],   # 转向时偏离
    ])
    
    output = model(current_gaze, vehicle_state)
    
    print("=== 正常驾驶 ===")
    print(f"未来视线轨迹: {output['future_gaze'][0]}")
    print(f"注意力区域概率: {output['attention_probs'][0]}")
    print(f"预测偏离: {output['predicted_deviation'][0]:.3f}")
    
    print("\n=== 分心驾驶 ===")
    print(f"未来视线轨迹: {output['future_gaze'][1]}")
    print(f"注意力区域概率: {output['attention_probs'][1]}")
    print(f"预测偏离: {output['predicted_deviation'][1]:.3f}")

---

Euro NCAP 2026 应用场景

1. ADAS 协同警告

ADAS 场景	视线预测应用	预期效果
前车碰撞预警	预测驾驶员是否将看向前方	优化警告时机
车道偏离预警	预测驾驶员注意力方向	提前警告
盲区检测	预测驾驶员是否将检查盲区	辅助提醒
自适应巡航	预测驾驶员对速度变化的注意力	调整策略

2. 分心趋势检测

class DistractionTrendDetector:
    """
    分心趋势检测器
    
    基于 ICCV 2025 论文方法
    
    关键：
    1. 检测视线偏离趋势（而非当前偏离）
    2. 提前 2-5 秒预测分心
    3. 与 ADAS 协同警告
    """
    
    def __init__(self,
                 prediction_horizon: int = 5,  # 预测未来 5 帧（约 1 秒）
                 trend_threshold: float = 0.3):  # 偏离趋势阈值
        self.predictor = GazeMovementPredictor(future_steps=prediction_horizon)
        self.trend_threshold = trend_threshold
        
        # 历史记录
        self.gaze_history = []
    
    def update(self, 
               current_gaze: np.ndarray,
               vehicle_state: np.ndarray) -> dict:
        """
        更新并预测
        
        Args:
            current_gaze: 当前视线 (pitch, yaw)
            vehicle_state: 车辆状态
        
        Returns:
            dict: {
                "trend_type": str,  # "stable", "drifting", "returning"
                "predicted_distraction": bool,
                "confidence": float,
                "recommended_action": str
            }
        """
        # 记录历史
        self.gaze_history.append(current_gaze)
        if len(self.gaze_history) > 30:
            self.gaze_history.pop(0)
        
        # 预测未来视线
        with torch.no_grad():
            output = self.predictor(
                torch.from_numpy(current_gaze).float().unsqueeze(0),
                torch.from_numpy(vehicle_state).float().unsqueeze(0)
            )
        
        future_gaze = output['future_gaze'][0].numpy()
        predicted_deviation = output['predicted_deviation'][0].item()
        
        # 趋势判断
        trend_type = self._classify_trend(future_gaze)
        
        # 分心预测
        predicted_distraction = predicted_deviation > self.trend_threshold
        
        # 置信度
        confidence = min(predicted_deviation / self.trend_threshold, 1.0)
        
        # 建议动作
        recommended_action = self._recommend_action(trend_type, predicted_distraction)
        
        return {
            "trend_type": trend_type,
            "predicted_distraction": predicted_distraction,
            "confidence": confidence,
            "predicted_deviation": predicted_deviation,
            "recommended_action": recommended_action
        }
    
    def _classify_trend(self, future_gaze: np.ndarray) -> str:
        """
        分类趋势
        
        - stable: 视线稳定（保持在道路中心附近）
        - drifting: 视线偏离趋势（逐渐偏离）
        - returning: 视线回归趋势（从偏离回归）
        """
        if len(self.gaze_history) < 10:
            return "stable"
        
        # 当前偏离
        current_deviation = np.linalg.norm(self.gaze_history[-1])
        
        # 未来偏离趋势
        future_deviations = np.linalg.norm(future_gaze, axis=1)
        
        # 判断趋势
        if future_deviations[-1] > current_deviation:
            return "drifting"  # 偏离增加
        elif future_deviations[-1] < current_deviation:
            return "returning"  # 偏离减少
        else:
            return "stable"  # 稳定
    
    def _recommend_action(self, 
                          trend_type: str,
                          predicted_distraction: bool) -> str:
        """建议动作"""
        if trend_type == "drifting" and predicted_distraction:
            return "提前警告 + ADAS 协同"
        elif trend_type == "drifting":
            return "关注趋势"
        elif trend_type == "returning":
            return "无需干预"
        else:
            return "正常监控"


# 测试示例
if __main__ == "__main__":
    detector = DistractionTrendDetector()
    
    # 模拟正常驾驶
    normal_gaze = np.array([0.05, 0.02])
    normal_vehicle = np.array([60, 0.1, 0.02, 0, 0, 0])
    
    normal_result = detector.update(normal_gaze, normal_vehicle)
    print("=== 正常驾驶 ===")
    print(f"趋势类型: {normal_result['trend_type']}")
    print(f"预测分心: {normal_result['predicted_distraction']}")
    print(f"建议动作: {normal_result['recommended_action']}")
    
    # 模拟分心趋势
    drifting_gaze = np.array([0.3, -0.2])
    drifting_vehicle = np.array([60, 0.3, 0.1, 0, 0, 0])
    
    drifting_result = detector.update(drifting_gaze, drifting_vehicle)
    print("\n=== 分心趋势 ===")
    print(f"趋势类型: {drifting_result['trend_type']}")
    print(f"预测分心: {drifting_result['predicted_distraction']}")
    print(f"建议动作: {drifting_result['recommended_action']}")

---

IMS 开发启示

1. 视线预测集成

graph TB
    A[当前视线] --> D[融合预测]
    B[车辆状态] --> D
    C[场景信息] --> D
    
    D --> E[未来视线轨迹]
    E --> F{趋势判断}
    
    F -->|drifting| G[提前警告]
    F -->|stable| H[正常监控]
    F -->|returning| I[无需干预]
    
    G --> J[ADAS 协同]

2. 与 Euro NCAP DSM 协同

DSM 场景	视线预测增强	预期改进
D-01: 手机使用	预测视线回归速度	优化警告取消时机
D-05: 视线偏离	预测持续偏离时间	提前检测
D-07: 疲劳特征	预测视线稳定性	疲劳趋势检测

---

参考资料

1. ICCV 2025 Paper PDF
2. GAZE Workshop 2026
3. Euro NCAP 2026 DSM Distraction Protocol
4. WACV 2024: Rotation-constrained Cross-view Gaze Estimation

---

总结

ICCV 2025 论文核心：

1. 视线-运动联合建模：预测未来注意力
2. 趋势检测：提前预测分心趋势
3. ADAS 协同：优化警告时机

IMS 开发优先级：

• 🔴 高：GazeMovementPredictor 实现
• 🟡 中：ADAS 协同接口开发
• 🟢 低：场景自适应优化

下一步行动：

• 实现 GazeMovementPredictor 模型
• 开发趋势检测逻辑
• 对齐 Euro NCAP D-05 场景