Early detection of occluded moving vehicles at intersections can prevent collisions, but most of the existing sensors only detect objects in sight. Sound propagates around obstacles by means such as reflection and diffraction, allowing passive acoustic sensing of the occluded vehicle. We propose a deep learning-based acoustic non-line-of-sight (NLOS) vehicle detection method. With the direction-of-arrival feature and time-frequency feature calculated from the microphone array data as inputs in image form, we designed and trained the parallel neural network to perceive the direction of the occluded moving vehicle at intersections. Since intelligent vehicles react differently to the approaching and leaving occluded moving vehicle, we further distinguished the occluded moving vehicle’s approaching/leaving status. To evaluate the proposed method, we collected data from different locations in the urban environment. The experimental results show that the classification for 6-class intersection traffic conditions reached 96.71%, and the occluded approaching vehicle was detected 1 second before it entered the line of sight, providing additional reaction time for intelligent vehicles. The direction of the occluded moving vehicle is accurately predicted, and the approaching/leaving status is further inferred, providing detailed traffic information for the intelligent vehicles’ response decisions. Furthermore, experiments show that the predictions of our method outperform the state-of-the-art acoustic NLOS approach vehicle detection baseline on real-world traffic datasets. Our code and dataset: https://github.com/RST2detection/Acoustic-Occluded-Vehicle-Detection.
1 January 2016 in Journal of Mechanical Engineering
The convolutional neural network (CNN) has achieved good performance for gas pipeline leak detection. However, the high computing cost hinders its application in embedded devices with limited hardware performance. To address this limitation, a model compression method called PWK is proposed, consisting of two parts: a Pruning-optimizing iterative strategy and Weight sharing combined with Knowledge distillation (KD). The pruning-optimizing iterative strategy consists of pruning and training phases. In the pruning phase, the compact network is trimmed using a mask derived from virtually pruning the original network, aiming to obtain the original network’s essential structure. In the training phase, important features extracted from the original network by the mask are transferred to the compact network through feature distillation to form a pruned network with low redundancy. KD was introduced into weight sharing, which can reduce weight storage to alleviate performance degradation. Finally, PWK compresses spectral enhancement CNN (SE-CNN) to an efficient CNN (SE-CNN-Lite) for gas pipeline leak detection. Experimental results show that the pruning-optimizing iterative strategy proposed in this paper has higher performance than other state-of-the-art methods under similar compression ratios. Our SE-CNN-Lite has lower redundancy than other state-of-the-art lightweight networks and achieves the optimal detection accuracy for gas pipeline leaks in low signal-to-noise ratio environments.