Les systèmes de détection d’objets dans le cadre d’un environnement routier

Abstract

Object detection has recently become a crucial component of safety-critical perception tasks in autonomous driving. Advancements in object detection within Automated Driving Systems (ADS) have largely been driven by the success of the Convolutional Neural Network (CNN). However, supervised passive learning of a deep object detection model is computationally costly and requires a large amount of annotated data to cover the wide diversity of objects and scenarios in vehicular environments. This presents challenges due to visual similarity and variability of objects and labeling costs. Consequently, acquiring this data is a time-consuming and expensive process, often requiring domain experts to manually annotate high-quality bounding boxes. Moreover, ensuring the functional safety of ADS requires robust object detectors, especially in critical situations encountered within this environment. In this context, the primary challenge is how to efficiently achieve the desired performance with a small set of labeled data while carefully balancing the trade-off between cost and accuracy. To deal with these limits, this thesis proposes and designs two contributions for object detection in autonomous driving, considering their characteristics and challenges: a batch-based query strategy and a Cost-Effective Deep Batch Mode Active Learning (CEDBMAL) framework. These solutions aim to ensure a robust and high-performance CNN-based object detector with low false detection rates and reduced annotation and training costs. Unlike the single-criterion query strategy, which may pick redundant or outlier samples that negatively affect the detector’s performance, our proposed batch-based query strategy automatically selects a batch of the top-ranked samples based on uncertainty and diversity criteria. These samples are more representative and informative, leading to more efficient training of the detector. To tailor classification uncertainty heuristics for deep object detection, we propose incorporating the detection model’s outputs, such as classification and regression predictions, into the uncertainty metric measurement, arguing that samples inducing uncertainty in the model are typically not outliers, but rather instances that exhibit a larger object distribution and are expected to improve its performance. To reduce the annotation burden from redundant samples, we also propose using Euclidean distance as a representativeness measure, quantifying diversity in termsof similarity. To mitigate the limitations posed by impractical batch size settings, our second proposal (CEDBMAL) combines our proposed query with labeling time prediction to design a cost-aware batch query. Initially, a set of batches with varied sizes is selected by the proposed query. Subsequently, we propose using labeling time prediction and dynamic programming to choose the best batch size by solving a 0-1 Knapsack problem under the constraints of annotation time, dataset size, and desired performance. This iterative process leads to an adaptive selection of the most useful and diverse training samples based on the cost of labeling. As a result, it becomes possible to effectively handle variations in annotation costs and significantly reduce the training and labeling expenses, both at the individual instance and batch level. To validate our approaches, extensive experiments were conducted on the Caltech Pedestrian dataset to fine-tune a pre-trained deep object detector (Tiny-YOLOv3) for pedestrian detection tasks. The effects of classification uncertainty, regression uncertainty, score aggregation methods, and batch size during sample selection were thoroughly investigated. The experimental results demonstrated that the uniform-cost DAL based on our proposed batch-based query strategy outperformed random sampling and transfer learning baselines. Moreover, our cost-effective DAL approach further boosted the performance compared to other baseline deep pedestrian detectors and uniform-cost DAL approaches with a specific deep pedestrian detector. Notably, our approaches enabled the development of a robust deep pedestrian detector with significantly fewer parameters, making it suitable for deployment on low-resource devices, while maintaining the detection error rate below 57%, saving up to 50% of the labeling effort, increasing the number of pedestrians detected at early cycles, and alleviating batch size dependency.