An Empirical Evaluation of Deep Learning on Highway Driving

TL;DR

The paper tackles highway perception by evaluating CNN-based lane and vehicle detection using a large, multi-sensor highway data collection. It adapts the Overfeat framework with a mask-detector to achieve real-time, single-pass detections, and extends the system to 3D lane boundary regression trained with perspective distortions. The study demonstrates real-time performance on GPUs, reports thorough lane and vehicle detection metrics, and validates depth predictions against radar ground truth. Overall, it provides evidence that deep learning can offer robust, real-time highway perception and outlines future work to leverage temporal information across frames.

Abstract

Numerous groups have applied a variety of deep learning techniques to computer vision problems in highway perception scenarios. In this paper, we presented a number of empirical evaluations of recent deep learning advances. Computer vision, combined with deep learning, has the potential to bring about a relatively inexpensive, robust solution to autonomous driving. To prepare deep learning for industry uptake and practical applications, neural networks will require large data sets that represent all possible driving environments and scenarios. We collect a large data set of highway data and apply deep learning and computer vision algorithms to problems such as car and lane detection. We show how existing convolutional neural networks (CNNs) can be used to perform lane and vehicle detection while running at frame rates required for a real-time system. Our results lend credence to the hypothesis that deep learning holds promise for autonomous driving.

Figures (12)

• Figure 1: Sample output from our neural network capable of lane and vehicle detection.
• Figure 2: mask detector
• Figure 3:
• Figure 4: Example of lane boundary ground truth
• Figure 5: Example output of lane detector after DBSCAN clustering