Accurate behavior monitoring of dairy cows is essential for enhancing animal welfare, farm productivity, and early detection of health issues, particularly in indoor (in-house) cattle environments. This study evaluates the performance of four YOLO-based object detection models: YOLOv5m, YOLOv5m with Convolutional Block Attention Module (CBAM), YOLOv7, and YOLOv8n for detecting key cow behaviors such as lying, standing, and feeding. The models were tested under three different lighting conditions (daylight, low-light, and mixed lighting) and two camera views (overhead and front). Results show that YOLOv7 achieved the highest detection accuracy, particularly in well-lit environments, while YOLOv5m+CBAM demonstrated superior performance under challenging lighting and occluded conditions. The overhead camera placement consistently yielded better performance compared to front view, due to improved posture visibility and fewer occlusions. Additionally, the inclusion of CBAM significantly improved detection performance, especially under front-view and low-light scenarios. These findings provide practical insights into deploying vision-based behavior monitoring systems in real-world dairy farm environments.