Optimization of Speed vs. Accuracy Trade-off in State-of-the-Art Object Detectors for Traffic Light Detection

Sammanfattning: Traffic lights detection systems are an important area of research, aimed towards improving the accuracy and response time of self-driving vehicles when faced with traffic signals. This project attempted to find a solution for the speed-accuracy trade-off faced by traffic light detection systems. In the study, a total of four STATE OF THE ART(SOAT) lights detection algorithms were used to compare the accuracy and speed of the detection and classification of traffic lights. The algorithms used include Faster RCNN + Inception V2, R-FCN + Resnet 101, SSD + Mobilenet V1, and YOLO v3 (416 x 416). All of these algorithms were tested on two datasets: BOSCH Small Traffic Lights Dataset and LISA Dataset. The results of these algorithms were analyzed through the measure of mAP @0.5. The speed of detection was measured in milliseconds. According to the results, FASTER RCNN + Inception V2 provided the most optimal speed-accuracy tradeoff score with a mAP value of 0.501 and 0.681 for BOSCH and LISA datasets respectively, along with 950.72 and 812.67 milliseconds detection time. All the other algorithms fared worse in accuracy while their gains in detection times were not enough to offset the improvement shown by Faster RCNN + Inception V2. Further research is needed to further study this tradeoff with different algorithm choices and replicate the results on more diverse and suitable datasets.