In Automated Driving Systems (ADS), the function of detecting objects on the road assists vehicle traffic and improves road safety. Deep Active Learning (DAL) is an advanced training method suitable for building robust Convolutional Neural Network (CNN)-based on road object detection models. This method automatically selects and manually labels training samples that are significantly less noisy, non-redundant and more useful. Depending on the complexity of detection task and the characteristics of urban scenes, the batch selection in the conventional batch mode DAL can suffer from the impact of the correlation between frame labels and batch size as well as variable labeling costs. This paper introduces a novel cost-effective-based training approach suitable for CNN-based on-road object detector, where frames labeling and batch size are considered in the sample selection process. We propose a batch sampling strategy that leverages the model prediction uncertainty along with dynamic programming to alleviate the selection batch size issue. Additionally, we investigate the effects of classification uncertainty, regression uncertainty and batch size during sample selection. Our approach was extensively validated on the Caltech Pedestrian dataset to fine-tune a pre-trained Tiny-YOLOv3 for performing pedestrian detection task. Results showed that our approach, compared to basic methods, can build robust detection model that keeps the detection error less than 57%, saving up 50% of the labeling effort and alleviating batch size dependency.