Optimal Sensor Data Fusion Architecture for Object Detection in Adverse Weather Conditions

TL;DR

Different sensor fusion architectures are compared for good and adverse weather conditions for finding the optimal fusion architecture for diverse weather situations and a new training strategy is introduced such that the performance of the object detector is greatly enhanced in adverse weather scenarios or if a sensor fails.

Abstract

A good and robust sensor data fusion in diverse weather conditions is a quite challenging task. There are several fusion architectures in the literature, e.g. the sensor data can be fused right at the beginning (Early Fusion), or they can be first processed separately and then concatenated later (Late Fusion). In this work, different fusion architectures are compared and evaluated by means of object detection tasks, in which the goal is to recognize and localize predefined objects in a stream of data. Usually, state-of-the-art object detectors based on neural networks are highly optimized for good weather conditions, since the well-known benchmarks only consist of sensor data recorded in optimal weather conditions. Therefore, the performance of these approaches decreases enormously or even fails in adverse weather conditions. In this work, different sensor fusion architectures are compared for good and adverse weather conditions for finding the optimal fusion architecture for diverse weather situations. A new training strategy is also introduced such that the performance of the object detector is greatly enhanced in adverse weather scenarios or if a sensor fails. Furthermore, the paper responds to the question if the detection accuracy can be increased further by providing the neural network with a-priori knowledge such as the spatial calibration of the sensors.

Figures (6)

• Figure 1: Result of a state-of-the-art object detector (Faster-RCNN Ren_2015_Faster_RCNN) if the lidar sensor fails.
• Figure 2: Result of the proposed object detector if the lidar sensor fails.
• Figure 3: Possible input features for the proposed object detector
• Figure 4: Early Fusion Approach: A common feature map is determined for all sensors by the feature encoder. For this, RGB image and the depth image determined from the lidar data are fit into a 4D tensor and progressed together in the neural network
• Figure 5: Late Fusion Approach: For each sensor, an independent feature map is determined by distinct feature encoder. The camera and lidar feature maps are then concatenated before the RPN is applied.