Depthwise Convolution For Compact Object Detector In nighttime Images

Abstract

Despite thermal imaging primarily used for nighttime surveillance, uniform temperature of object and background makes it difficult to acquire details in the scene being observed and thereby object detection. Further, thermal images collected over long distances degrade the spatial resolution of the acquired objects and so do the moving objects leading to noisy features. We present a computationally efficient object detection approach using Depthwise Deep Convolutional Neural Network (DDCNN) for detecting and classifying objects in nighttime images under low resolution. The Depthwise Convolution (DC) employed in the proposed approach minimises the network’s computational complexity resulting in the lowest number of training parameters (i.e., 3M) as compared to the other existing state-of-the-art methods such as FRCNN (52M), SSD (24M) and YOLO-v3 (61M) parameters. Further, by introducing novel Tversky and Intersection over Union (IoU) loss functions into the compact architectural design, we improve nighttime object detection accuracy. The validity of the proposed model is assessed on numerous datasets such as FLIR, KAIST, MS, and our internal dataset having multiple objects in each image. The experimental results from the proposed method indicate both quantitative and qualitative improvements over the recent state-of-the-art methods for nighttime imaging. The proposed approach achieves a mean Average Precision (mAP) of 52.39% and a highest individual object detection accuracy of 72.70% accuracy for cars in nigh-time situations suggesting applications in real-time use cases.