3D Object Detection for Self-Driving Cars Using Video and LiDAR: An Ablation Study.

Methods based on 64-beam LiDAR can provide very precise 3D object detection. However, highly accurate LiDAR sensors are extremely costly: a 64-beam model can cost approximately USD 75,000. We previously proposed SLS-Fusion (sparse LiDAR and stereo fusion) to fuse low-cost four-beam LiDAR with stereo cameras that outperform most advanced stereo-LiDAR fusion methods. In this paper, and according to the number of LiDAR beams used, we analyzed how the stereo and LiDAR sensors contributed to the performance of the SLS-Fusion model for 3D object detection. Data coming from the stereo camera play a significant role in the fusion model. However, it is necessary to quantify this contribution and identify the variations in such a contribution with respect to the number of LiDAR beams used inside the model. Thus, to evaluate the roles of the parts of the SLS-Fusion network that represent LiDAR and stereo camera architectures, we propose dividing the model into two independent decoder networks. The results of this study show that-starting from four beams-increasing the number of LiDAR beams has no significant impact on the SLS-Fusion performance. The presented results can guide the design decisions by practitioners.

3D Object Detection with SLS-Fusion Network in Foggy Weather Conditions.

The role of sensors such as cameras or LiDAR (Light Detection and Ranging) is crucial for the environmental awareness of self-driving cars. However, the data collected from these sensors are subject to distortions in extreme weather conditions such as fog, rain, and snow. This issue could lead to many safety problems while operating a self-driving vehicle. The purpose of this study is to analyze the effects of fog on the detection of objects in driving scenes and then to propose methods for improvement. Collecting and processing data in adverse weather conditions is often more difficult than data in good weather conditions. Hence, a synthetic dataset that can simulate bad weather conditions is a good choice to validate a method, as it is simpler and more economical, before working with a real dataset. In this paper, we apply fog synthesis on the public KITTI dataset to generate the Multifog KITTI dataset for both images and point clouds. In terms of processing tasks, we test our previous 3D object detector based on LiDAR and camera, named the Spare LiDAR Stereo Fusion Network (SLS-Fusion), to see how it is affected by foggy weather conditions. We propose to train using both the original dataset and the augmented dataset to improve performance in foggy weather conditions while keeping good performance under normal conditions. We conducted experiments on the KITTI and the proposed Multifog KITTI datasets which show that, before any improvement, performance is reduced by 42.67% in 3D object detection for Moderate objects in foggy weather conditions. By using a specific strategy of training, the results significantly improved by 26.72% and keep performing quite well on the original dataset with a drop only of 8.23%. In summary, fog often causes the failure of 3D detection on driving scenes. By additional training with the augmented dataset, we significantly improve the performance of the proposed 3D object detection algorithm for self-driving cars in foggy weather conditions.

A Unified Deep Framework for Joint 3D Pose Estimation and Action Recognition from a Single RGB Camera.

We present a deep learning-based multitask framework for joint 3D human pose estimation and action recognition from RGB sensors using simple cameras. The approach proceeds along two stages. In the first, a real-time 2D pose detector is run to determine the precise pixel location of important keypoints of the human body. A two-stream deep neural network is then designed and trained to map detected 2D keypoints into 3D poses. In the second stage, the Efficient Neural Architecture Search (ENAS) algorithm is deployed to find an optimal network architecture that is used for modeling the spatio-temporal evolution of the estimated 3D poses via an image-based intermediate representation and performing action recognition. Experiments on Human3.6M, MSR Action3D and SBU Kinect Interaction datasets verify the effectiveness of the proposed method on the targeted tasks. Moreover, we show that the method requires a low computational budget for training and inference. In particular, the experimental results show that by using a monocular RGB sensor, we can develop a 3D pose estimation and human action recognition approach that reaches the performance of RGB-depth sensors. This opens up many opportunities for leveraging RGB cameras (which are much cheaper than depth cameras and extensively deployed in private and public places) to build intelligent recognition systems.

Spatio⁻Temporal Image Representation of 3D Skeletal Movements for View-Invariant Action Recognition with Deep Convolutional Neural Networks.

Designing motion representations for 3D human action recognition from skeleton sequences is an important yet challenging task. An effective representation should be robust to noise, invariant to viewpoint changes and result in a good performance with low-computational demand. Two main challenges in this task include how to efficiently represent spatio-temporal patterns of skeletal movements and how to learn their discriminative features for classification tasks. This paper presents a novel skeleton-based representation and a deep learning framework for 3D action recognition using RGB-D sensors. We propose to build an action map called SPMF (Skeleton Posture-Motion Feature), which is a compact image representation built from skeleton poses and their motions. An Adaptive Histogram Equalization (AHE) algorithm is then applied on the SPMF to enhance their local patterns and form an enhanced action map, namely Enhanced-SPMF. For learning and classification tasks, we exploit Deep Convolutional Neural Networks based on the DenseNet architecture to learn directly an end-to-end mapping between input skeleton sequences and their action labels via the Enhanced-SPMFs. The proposed method is evaluated on four challenging benchmark datasets, including both individual actions, interactions, multiview and large-scale datasets. The experimental results demonstrate that the proposed method outperforms previous state-of-the-art approaches on all benchmark tasks, whilst requiring low computational time for training and inference.