Real-time vehicle detection using low-cost sensors
2019-11-26T16:08:13Z (GMT) by
Improving road safety and reducing the number of accidents is one of the top priorities for the automotive industry. As human driving behaviour is one of the top causation factors of road accidents, research is working towards removing control from the human driver by automating functions and finally introducing a fully Autonomous Vehicle (AV). A Collision Avoidance System (CAS) is one of the key safety systems for an AV, as it ensures all potential threats ahead of the vehicle are identified and appropriate action is taken. This research focuses on the task of vehicle detection, which is the base of a CAS, and attempts to produce an effective vehicle detector based on the data coming from a low-cost monocular camera. Developing a robust CAS based on low-cost sensor is crucial to bringing the cost of safety systems down and in this way, increase their adoption rate by end users. In this work, detectors are developed based on the two main approaches to vehicle detection using a monocular camera. The first is the traditional image processing approach where visual cues are utilised to generate potential vehicle locations and at a second stage, verify the existence of vehicles in an image. The second approach is based on a Convolutional Neural Network, a computationally expensive method that unifies the detection process in a single pipeline. The goal is to determine which method is more appropriate for real-time applications. Following the first approach, a vehicle detector based on the combination of HOG features and SVM classification is developed. The detector attempts to optimise performance by modifying the detection pipeline and improve run-time performance. For the CNN-based approach, six different network models are developed and trained end to end using collected data, each with a different network structure and parameters, in an attempt to determine which combination produces the best results. The evaluation of the different vehicle detectors produced some interesting findings; the first approach did not manage to produce a working detector, while the CNN-based approach produced a high performing vehicle detector with an 85.87% average precision and a very low miss rate. The detector managed to perform well under different operational environments (motorway, urban and rural roads) and the results were validated using an external dataset. Additional testing of the vehicle detector indicated it is suitable as a base for safety applications such as CAS, with a run time performance of 12FPS and potential for further improvements.