A Multiobject Detection Scheme Based on Deep Learning for Infrared Images

Abstract

Multi-object detection is one of essential prerequisites for autonomous driving under complicated traffic environments. Specially, for the cases of nighttime driving, the accuracy of object detector for those images captured by infrared cameras is inevitably affected by several unanticipated issues (such as dim and unexpected light sources), which may significantly degrade the performance of detecting multiple objects (e.g., pedestrians and vehicles). To address such problem, we propose a <i>Light-Weighted Multi-Object Detection Scheme for Infrared Images</i> (<i>LW-MODS-IRI</i>). First, we select an effective super-resolution (SR) method to enhance and repair those infrared images, which can facilitate restoring the details of images for better detection accuracy. Next, those enhanced images are trained by our LW-MODS-IRI scheme. In contrast with previous methods, we develop an adaptive computing policy for anchor clustering that aims at locating the bound of targets in a reasonable manner for higher detection precision. Moreover, a modified attention mechanism is exploited to enlarge the receptive fields for promoting the detection performance of those large objects. Furthermore, we propose an improved channel-level pruning technique to generate both compact and accurate model after the training with sparsity. Finally, the extensive experiment results on Teledyne FLIR ADAS dataset (FLIR) demonstrate that the LW-MODS-IRI scheme can have both higher inference accuracy (e.g., 10.7% more) and higher detection efficiency (e.g., 4.13 times faster) compared to the existing object detectors, which typically reveals its <i>practical viability</i> when it is applied to in-vehicle systems.