Binocular vision vehicle detection method based on improved Fast-RCNN

Vehicles with different spatial scales exhibit significantly different characteristics, resulting in low efficiency, poor accuracy of vehicle detection methods and difficulty for accurately obtaining vehicle distance information. To solve this problem, a vehicle detection method based on improved fast region-based convolutional neural network (Fast-RCNN) to detect vehicle targets was proposed, which using binocular vision to range vehicles. Firstly the binocular stereo camera was used to acquire the image information and perform preprocessing, the training data of the deep neural network Fast-RCNN was loaded, then multiple built-in sub networks were introduced for different spatial scales of vehicles, and the output of all sub networks was adaptively combined to detect vehicles. At last the speeded up robust features (SURF) matching algorithm was utilized to carry out the stereo matching of the left and right images.And based on the matching data, the 3D reconstruction and vehicle centroid coordinates were determined so as to measure the distance between vehicle and binocular camera. Experimental results show that the algorithm can achieve fast detection of vehicles, the detection time is shorter than the traditional Fast-RCNN by 42 ms and the accurate measurement of vehicle distance in 5 m can be achieved with an error of only 2.4%. The proposed method has high accuracy and good real-time performance.