Adaptive Feature-Manipulated Vehicle and Pedestrian Detection in Infrared Images

Infrared imaging is widely applied in assisted driving systems to enable night vision for sensitive targets such as vehicles and pedestrians. However, the detection accuracy of these targets is limited on lightweight detection networks due to their coarse color and texture characteristics rendered in infrared images. To solve this problem, we propose an adaptive feature-manipulated network (AFMNet) for accurate vehicle and pedestrian detection in infrared images. First, a refined spatial pooling module that uses one-dimensional convolution is proposed to establish local and channel feature mapping under different receptive fields so that different fine-grained features are fused. Second, the shuffle manipulation module is designed which includes slicing and shuffling to manipulate the spatial and channel features so that the information loss problem caused by conventional convolutional downsampling is overcome. Third, the adaptive connection of features at different scales using learnable parameters is proposed, and then the target features are reinforced by location and channel calibration branches. The experimental results show that AFMNet achieves the best performance in terms of average detection accuracy of 86.4%, model size of 5.3MB, and detection speed of 48FPS on GTX 2080Ti.