智能识别不礼让行人行为算法模型的图像训练数据
收藏资源简介：
"本数据集主要用于提升AI模型对机动车斑马线未礼让行人违法行为的识别能力与精确性。通过对该数据集的训练，使AI模型能够实现对违法行为的精准识别，并可应用于城市道路、学校区域、商业区等各类斑马线场景的智能交通执法系统。同时，本数据集可为交通管理部门提供智能化执法依据，有效提升机动车礼让行人的执法效率，助力构建文明安全的道路交通环境。1.数据采集 通过企业自有摄像设备自行采集道路车辆图像，同步记录图像ID、采集时间、设备型号、地理坐标、光照条件、天气状况、信号灯状态等数据。 2.数据预处理与标注 通过数据清洗剔除模糊、重复图像。按6:2:2比例划分训练集/验证集/测试集。设置多级标注体系： 一级标签：礼让/未礼让 二级标签：车辆未减速/行人被迫避让（步态突变标记）/信号灯冲突（行人绿灯时车辆闯过） 辅助标注：车辆边界框坐标、行人边界框坐标 3.模型选择与初始化 采用YOLOv8s+HRNet-w32混合架构，初始化参数并优化超参数：批量大小1-16动态调整，学习率0.01-0.001动态调整，调整检测车辆锚框参数适配前脸形态，集成时空注意力机制建模车辆-行人交互关系。 4.模型训练 基于PyTorch实施分布式训练，采用混合精度训练（FP16）提升效率。数据增强涵盖夜间红外噪点模拟、雨雾行人遮挡及密集人群干扰生成，设置早停机制（patience=12）和梯度裁剪（max_norm=2.0）防止过拟合。 5.模型评估 在训练模型的过程中，使用验证集调整超参数，训练完成后在测试集上评估模型表现，评估指标包含： 基础性能：车辆检测mAP@0.5、误报率 场景鲁棒性测试：夜间检出率
This dataset is primarily designed to enhance the recognition capability and accuracy of AI models in identifying the traffic violation of motor vehicles failing to yield to pedestrians at zebra crossings. Through training on this dataset, AI models can achieve precise recognition of such violations, and can be applied to intelligent traffic law enforcement systems across various zebra crossing scenarios such as urban roads, school zones, and commercial areas. Additionally, this dataset can provide intelligent law enforcement references for traffic management departments, effectively improving the efficiency of law enforcement regarding motor vehicles yielding to pedestrians, and contributing to the construction of a civilized and safe road traffic environment. 1. Data Collection Road vehicle images were collected using the enterprise's own camera equipment, with synchronized recording of data including image ID, collection time, equipment model, geographic coordinates, lighting conditions, weather conditions, and traffic light status. 2. Data Preprocessing and Annotation Blurry and duplicate images were removed via data cleaning. The dataset was split into training, validation, and test sets at a ratio of 6:2:2. A multi-level annotation system was established: - Primary labels: Yield / Fail to yield - Secondary labels: Vehicle not decelerating / Pedestrians forced to evade (marked by sudden gait changes) / Signal light conflict (vehicles running red lights when pedestrians have green light) - Auxiliary annotations: Vehicle bounding box coordinates, pedestrian bounding box coordinates 3. Model Selection and Initialization A hybrid architecture of YOLOv8s + HRNet-w32 was adopted, with initialization of parameters and optimization of hyperparameters: dynamically adjusting batch size from 1 to 16, dynamically adjusting learning rate from 0.01 to 0.001, adjusting detection vehicle anchor box parameters to adapt to vehicle front-end shapes, and integrating spatial-temporal attention mechanisms to model vehicle-pedestrian interaction relationships. 4. Model Training Distributed training was implemented based on PyTorch, with mixed-precision training (FP16) adopted to improve efficiency. Data augmentation includes simulation of nighttime infrared noise, generation of rain-fog pedestrian occlusion and dense crowd interference. An early stopping mechanism (patience=12) and gradient clipping (max_norm=2.0) were set to prevent overfitting. 5. Model Evaluation During model training, the validation set was used to adjust hyperparameters. After training, model performance was evaluated on the test set, with evaluation metrics including: - Basic performance: mAP@0.5 for vehicle detection, false positive rate - Scene robustness test: Nighttime detection rate
